Viewing: lov_io.c
// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
* Copyright (c) 2011, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
*
* Implementation of cl_io for LOV layer.
*
* Author: Nikita Danilov <nikita.danilov@sun.com>
* Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
*/
#define DEBUG_SUBSYSTEM S_LOV
#include <linux/lustre/erasure_code.h>
#include "lov_cl_internal.h"
/** \addtogroup lov
* @{
*/
/**
* lov_sub_alloc() - Allocate a new sub IO
* @lio: top level lov IO structure
* @index: index into lov (stripe)
*
* Return Pointer to allocated lov_io_sub structure
*/
static inline struct lov_io_sub *lov_sub_alloc(struct lov_io *lio, int index)
{
struct lov_io_sub *sub;
if (lio->lis_nr_subios == 0) {
LASSERT(lio->lis_single_subio_index == -1);
sub = &lio->lis_single_subio;
lio->lis_single_subio_index = index;
memset(sub, 0, sizeof(*sub));
} else {
OBD_ALLOC_PTR(sub);
}
if (sub) {
INIT_LIST_HEAD(&sub->sub_list);
INIT_LIST_HEAD(&sub->sub_linkage);
sub->sub_subio_index = index;
}
return sub;
}
/**
* lov_sub_free() - Release a sub IO
* @lio: top level lov IO structure
* @sub: sub io to individual stripe
*/
static inline void lov_sub_free(struct lov_io *lio, struct lov_io_sub *sub)
{
if (sub->sub_subio_index == lio->lis_single_subio_index) {
LASSERT(sub == &lio->lis_single_subio);
lio->lis_single_subio_index = -1;
} else {
OBD_FREE_PTR(sub);
}
}
static void lov_io_sub_fini(const struct lu_env *env, struct lov_io *lio,
struct lov_io_sub *sub)
{
ENTRY;
cl_io_fini(sub->sub_env, &sub->sub_io);
if (sub->sub_env && !IS_ERR(sub->sub_env)) {
cl_env_put(sub->sub_env, &sub->sub_refcheck);
sub->sub_env = NULL;
}
EXIT;
}
static inline bool
is_index_within_mirror(struct lov_object *lov, int index, int mirror_index)
{
struct lov_mirror_entry *lre;
if (mirror_index < 0)
return false;
lre = lov_mirror_entry(lov, mirror_index);
return (index >= lre->lre_start && index <= lre->lre_end);
}
static int lov_io_sub_init(const struct lu_env *env, struct lov_io *lio,
struct lov_io_sub *sub)
{
struct lov_object *lov = lio->lis_object;
struct cl_io *sub_io;
struct cl_object *sub_obj;
struct cl_io *io = lio->lis_cl.cis_io;
int index = lov_comp_entry(sub->sub_subio_index);
int stripe = lov_comp_stripe(sub->sub_subio_index);
int result = 0;
LASSERT(sub->sub_env == NULL);
ENTRY;
if (unlikely(!lov_r0(lov, index)->lo_sub ||
!lov_r0(lov, index)->lo_sub[stripe]))
RETURN(-EIO);
LASSERTF(ergo(lov_is_flr(lov),
/* within data mirror */
is_index_within_mirror(lov, index,
lio->lis_mirror_index) ||
/* within parity mirror */
is_index_within_mirror(lov, index,
lov_parity_mirror_index_from_data(
lio,
lio->lis_mirror_index))),
DFID "iot = %d, index = %d, mirror = %d\n",
PFID(lu_object_fid(lov2lu(lov))), io->ci_type, index,
lio->lis_mirror_index);
/* obtain new environment */
sub->sub_env = cl_env_get(&sub->sub_refcheck);
if (IS_ERR(sub->sub_env)) {
result = PTR_ERR(sub->sub_env);
RETURN(result);
}
sub_obj = lovsub2cl(lov_r0(lov, index)->lo_sub[stripe]);
sub_io = &sub->sub_io;
sub_io->ci_obj = sub_obj;
sub_io->ci_result = 0;
sub_io->ci_parent = io;
sub_io->ci_lockreq = io->ci_lockreq;
sub_io->ci_type = io->ci_type;
sub_io->ci_no_srvlock = io->ci_no_srvlock;
sub_io->ci_noatime = io->ci_noatime;
sub_io->ci_async_readahead = io->ci_async_readahead;
sub_io->ci_lock_no_expand = io->ci_lock_no_expand;
sub_io->ci_ndelay = io->ci_ndelay;
sub_io->ci_layout_version = io->ci_layout_version;
sub_io->ci_tried_all_mirrors = io->ci_tried_all_mirrors;
sub_io->ci_parity_io = io->ci_parity_io;
sub_io->ci_parity_eof = io->ci_parity_eof;
result = cl_io_sub_init(sub->sub_env, sub_io, io->ci_type, sub_obj);
if (result < 0)
lov_io_sub_fini(env, lio, sub);
RETURN(result);
}
struct lov_io_sub *lov_sub_get(const struct lu_env *env,
struct lov_io *lio, int index)
{
struct lov_io_sub *sub;
int rc = 0;
ENTRY;
list_for_each_entry(sub, &lio->lis_subios, sub_list) {
if (sub->sub_subio_index == index) {
rc = 1;
break;
}
}
if (rc == 0) {
sub = lov_sub_alloc(lio, index);
if (!sub)
GOTO(out, rc = -ENOMEM);
rc = lov_io_sub_init(env, lio, sub);
if (rc < 0) {
lov_sub_free(lio, sub);
GOTO(out, rc);
}
list_add_tail(&sub->sub_list, &lio->lis_subios);
lio->lis_nr_subios++;
}
out:
if (rc < 0)
sub = ERR_PTR(rc);
else
sub->sub_io.ci_noquota = lio->lis_cl.cis_io->ci_noquota;
RETURN(sub);
}
/*
* Lov io operations.
*/
/**
* lov_io_subio_init() - Initilize LOV I/O operation
* @env: lustre environment
* @lio: Pointer to struct lov_io
* @io: highlevel I/O request
*
* Return:
* * %0 on success
* * %negative on failure
*/
static int lov_io_subio_init(const struct lu_env *env, struct lov_io *lio,
struct cl_io *io)
{
ENTRY;
LASSERT(lio->lis_object != NULL);
INIT_LIST_HEAD(&lio->lis_subios);
lio->lis_single_subio_index = -1;
lio->lis_nr_subios = 0;
RETURN(0);
}
/*
* Decide if it will need write intent RPC
*/
static int lov_io_mirror_write_intent(struct lov_io *lio,
struct lov_object *obj, struct cl_io *io)
{
struct lu_object *lobj = lov2lu(obj);
struct lov_device *dev = lov_object_dev(obj);
struct lov_layout_composite *comp = &obj->u.composite;
struct lu_extent *ext = &io->ci_write_intent;
struct lov_mirror_entry *lre;
struct lov_mirror_entry *primary;
struct lov_layout_entry *lle;
size_t count = 0;
ENTRY;
*ext = (typeof(*ext)) { lio->lis_pos, lio->lis_endpos };
io->ci_need_write_intent = 0;
if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
cl_io_is_fallocate(io) || cl_io_is_trunc(io)))
RETURN(0);
/*
* FLR: check if it needs to send a write intent RPC to server.
* Writing to sync_pending file needs write intent RPC to change
* the file state back to write_pending, so that the layout version
* can be increased when the state changes to sync_pending at a later
* time. Otherwise there exists a chance that an evicted client may
* dirty the file data while resync client is working on it.
* Designated I/O is allowed for resync workload.
*/
if (lov_flr_state(obj) == LCM_FL_RDONLY ||
(lov_flr_state(obj) == LCM_FL_SYNC_PENDING &&
io->ci_designated_mirror == 0)) {
io->ci_need_write_intent = 1;
RETURN(0);
}
LASSERT((lov_flr_state(obj) == LCM_FL_WRITE_PENDING));
LASSERT(comp->lo_preferred_mirror >= 0);
/*
* need to iterate all components to see if there are
* multiple components covering the writing component
*/
primary = &comp->lo_mirrors[comp->lo_preferred_mirror];
if (primary->lre_stale || !primary->lre_valid) {
/**
* new server could pick a primary mirror which old client
* does not recognize, and old client would mark it as
* invalid.
*/
CERROR("%s:"DFID": cannot find known valid non-stale mirror, could be new server picked a mirror which this client "
"does not recognize.\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lobj)));
RETURN(-EIO);
}
lov_foreach_mirror_layout_entry(obj, lle, primary) {
LASSERT(lle->lle_valid);
if (!lu_extent_is_overlapped(ext, lle->lle_extent))
continue;
ext->e_start = min(ext->e_start, lle->lle_extent->e_start);
ext->e_end = max(ext->e_end, lle->lle_extent->e_end);
++count;
}
if (count == 0) {
CERROR("%s:"DFID": cannot find any valid components covering file extent "
DEXT", mirror: %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lobj)), PEXT(ext),
primary->lre_mirror_id);
RETURN(-EIO);
}
count = 0;
lov_foreach_mirror_entry(obj, lre) {
if (lre == primary)
continue;
lov_foreach_mirror_layout_entry(obj, lle, lre) {
if (!lle->lle_valid)
continue;
if (lu_extent_is_overlapped(ext, lle->lle_extent)) {
++count;
break;
}
}
}
CDEBUG(D_VFSTRACE, DFID "there are %zd components to be staled to modify file extent "
DEXT", iot: %d\n",
PFID(lu_object_fid(lobj)), count, PEXT(ext), io->ci_type);
io->ci_need_write_intent = count > 0;
RETURN(0);
}
static bool lov_find_comp_with_ec(struct lov_io *lio, struct lov_object *lov,
int *mirror_index)
{
struct lov_device *dev = lov_object_dev(lov);
struct lov_layout_composite *comp = &lov->u.composite;
struct lu_extent ext = { .e_start = lio->lis_pos,
.e_end = lio->lis_endpos };
unsigned short mirror_count;
__u16 data_mirror_id;
int i;
ENTRY;
/* EC components are only used in FLR */
if (!lov_is_flr(lov))
RETURN(false);
mirror_count = comp->lo_mirror_count;
for (i = 0; i < mirror_count; i++) {
struct lov_layout_entry *parity_lle = NULL;
struct lov_layout_entry *lle;
lov_foreach_mirror_layout_entry(lov, lle,
&comp->lo_mirrors[i]) {
if (!(lle->lle_lsme->lsme_flags & LCME_FL_PARITY))
continue;
/* parity component */
if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
if (lle->lle_lsme->lsme_flags & LCME_FL_STALE) {
/* skip stale parity component */
parity_lle = NULL;
break;
}
/*
* capture the match while 'lle' is valid --
* after the loop it points one past the last
* entry (lle_lsme == NULL).
*/
parity_lle = lle;
}
}
if (parity_lle) {
/*
* a parity component's mirror_link_id is the id of
* the data mirror it protects.
*/
data_mirror_id =
parity_lle->lle_lsme->lsme_mirror_link_id;
*mirror_index = lov_mirror_index_by_id(
lov, data_mirror_id);
if (*mirror_index < 0) {
CERROR("%s:"DFID": failed to find data mirror %u: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))),
data_mirror_id, *mirror_index);
RETURN(false);
}
RETURN(true);
}
}
RETURN(false);
}
/**
* lov_io_parity_size_comp() - Parity byte span for one data/parity pair
* @data_lsme: data component stripe metadata
* @parity_lsme: parity component stripe metadata
* @data_end: exclusive end of data in file offset space
*
* Returns how many bytes of parity are written for data in
* [data_lsme->lsme_extent.e_start, data_end).
* May overestimate (holes) but must not be short.
*/
static loff_t lov_io_parity_size_comp(struct lov_stripe_md_entry *data_lsme,
struct lov_stripe_md_entry *parity_lsme,
loff_t data_end)
{
struct ec_split_comp sc;
loff_t comp_start = data_lsme->lsme_extent.e_start;
loff_t data_upto;
__u64 stripe_set_size;
__u64 num_stripe_sets;
__u64 data_in_extent;
int num_raid_sets;
if (data_end <= comp_start)
return 0;
data_upto = min_t(__u64, (__u64)data_end,
data_lsme->lsme_extent.e_end);
ec_split_stripes(data_lsme->lsme_stripe_count,
parity_lsme->lsme_dstripe_count, &sc);
num_raid_sets = sc.esc_n0 + sc.esc_n1;
stripe_set_size = (__u64)data_lsme->lsme_stripe_count *
data_lsme->lsme_stripe_size;
data_in_extent = data_upto - comp_start;
num_stripe_sets = (data_in_extent + stripe_set_size - 1) /
stripe_set_size;
return num_stripe_sets * num_raid_sets *
parity_lsme->lsme_cstripe_count *
parity_lsme->lsme_stripe_size;
}
/**
* lov_io_find_data_comp() - Find data component matching parity extent
* @obj: lov object
* @data_lre: data mirror entry
* @parity_ext: parity component extent
*
* Return: matching layout entry, or NULL
*/
static struct lov_layout_entry *
lov_io_find_data_comp(struct lov_object *obj, struct lov_mirror_entry *data_lre,
struct lu_extent *parity_ext)
{
struct lov_layout_entry *data_lle;
lov_foreach_mirror_layout_entry(obj, data_lle, data_lre) {
if (lu_extent_is_equal(parity_ext,
&data_lle->lle_lsme->lsme_extent))
return data_lle;
}
return NULL;
}
/**
* lov_io_parity_eof() - Parity mirror EOF from @data_size and RAID geometry
* @obj: lov object
* @mirror_index: parity mirror index
* @data_size: data mirror file size
*
* For each parity component below @data_size, add lov_io_parity_size_comp() to
* its extent start; return the maximum. May overestimate but not underestimate.
*
* Return: parity EOF in file offset space, or 0 on error
*/
static loff_t lov_io_parity_eof(struct lov_object *obj, int mirror_index,
loff_t data_size)
{
struct lov_device *dev = lov_object_dev(obj);
struct lov_mirror_entry *parity_lre;
struct lov_mirror_entry *data_lre;
struct lov_layout_entry *parity_lle;
struct lov_layout_entry *data_lle;
struct lu_extent *ext;
__u16 data_mirror_id;
loff_t parity_eof = 0;
if (data_size == 0)
return 0;
parity_lre = lov_mirror_entry(obj, mirror_index);
parity_lle = lov_entry(obj, parity_lre->lre_start);
if (!(parity_lle->lle_lsme->lsme_pattern & LOV_PATTERN_PARITY))
return 0;
data_mirror_id = parity_lle->lle_lsme->lsme_mirror_link_id;
data_lre = lov_mirror_by_id(obj, data_mirror_id);
if (data_lre == NULL)
return 0;
lov_foreach_mirror_layout_entry(obj, parity_lle, parity_lre) {
loff_t seg_eof;
ext = &parity_lle->lle_lsme->lsme_extent;
if (data_size <= ext->e_start)
continue;
data_lle = lov_io_find_data_comp(obj, data_lre, ext);
if (data_lle == NULL) {
CERROR("%s:"DFID": no data component matching parity extent "DEXT" (parity mirror %u, data mirror %u)\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(obj))), PEXT(ext),
parity_lre->lre_mirror_id, data_mirror_id);
continue;
}
/*
* Or we could just find the last valid data component,
* e.g. the one with actual data written? For security, I
* prefer not to do this here.
*/
seg_eof = ext->e_start +
lov_io_parity_size_comp(data_lle->lle_lsme,
parity_lle->lle_lsme,
data_size);
if (seg_eof > parity_eof)
parity_eof = seg_eof;
}
CDEBUG(D_INODE, DFID ": parity EOF calc: data_size=%lld parity_eof=%lld\n",
PFID(lu_object_fid(lov2lu(obj))), data_size, parity_eof);
return parity_eof;
}
static loff_t lov_io_top_file_size(const struct lu_env *env,
struct lov_object *obj)
{
struct cl_object *top_obj = cl_object_top(&obj->lo_cl);
struct cl_attr attr = { 0 };
cl_object_attr_lock(top_obj);
if (top_obj->co_ops->coo_attr_get)
top_obj->co_ops->coo_attr_get(env, top_obj, &attr);
cl_object_attr_unlock(top_obj);
return attr.cat_size;
}
static int lov_io_mirror_init(const struct lu_env *env, struct lov_io *lio,
struct lov_object *obj, struct cl_io *io)
{
struct lov_device *dev = lov_object_dev(obj);
struct lov_layout_composite *comp = &obj->u.composite;
bool skipped_parity = false;
int index = -ENOENT;
int i;
int result = 0;
ENTRY;
if (io->ci_type == CIT_EC_RD) {
if (lov_find_comp_with_ec(lio, obj, &index)) {
lio->lis_mirror_index = index;
GOTO(ndelay_retry, result);
} else {
result = -EIO;
CWARN("%s: "DFID": no updated parity component to restore data: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(obj))), result);
RETURN(result);
}
}
if (!lov_is_flr(obj)) {
/* only locks/pages are manipulated for CIT_MISC op, no
* cl_io_loop() will be called, don't check/set mirror info.
*/
if (io->ci_type != CIT_MISC) {
LASSERT(comp->lo_preferred_mirror == 0);
lio->lis_mirror_index = comp->lo_preferred_mirror;
}
io->ci_ndelay = 0;
RETURN(0);
}
/* transfer the layout version for verification */
if (io->ci_layout_version == 0)
io->ci_layout_version = obj->lo_lsm->lsm_layout_gen;
/* find the corresponding mirror for designated mirror IO */
if (io->ci_designated_mirror > 0) {
struct lov_mirror_entry *entry;
LASSERT(!io->ci_ndelay);
CDEBUG(D_LAYOUT, "designated I/O mirror state: %d\n",
lov_flr_state(obj));
if ((cl_io_is_trunc(io) || io->ci_type == CIT_WRITE ||
cl_io_is_fallocate(io)) &&
(io->ci_layout_version != obj->lo_lsm->lsm_layout_gen)) {
/*
* For resync I/O, the ci_layout_version was the layout
* version when resync starts. If it doesn't match the
* current object layout version, it means the layout
* has been changed
*/
RETURN(-ESTALE);
}
io->ci_layout_version |= LU_LAYOUT_RESYNC;
index = 0;
lio->lis_mirror_index = -1;
lov_foreach_mirror_entry(obj, entry) {
if (entry->lre_mirror_id ==
io->ci_designated_mirror) {
lio->lis_mirror_index = index;
break;
}
index++;
}
if (lio->lis_mirror_index < 0)
RETURN(-EINVAL);
/* Set ci_parity_io if this is a parity mirror */
io->ci_parity_io =
lov_mirror_entry(obj, lio->lis_mirror_index)->lre_parity;
/* Calculate parity EOF if this is a parity mirror */
if (io->ci_parity_io) {
loff_t file_size;
/*
* Get file size from the top (VVP) layer only.
* We can't use cl_object_attr_get() here because it
* walks through all layers and the LOV layer would
* reset cat_size to 0 before recalculating from
* stripe attributes (which may not be populated yet).
* The VVP layer's coo_attr_get reads i_size directly
* from the inode.
*/
file_size = lov_io_top_file_size(env, obj);
io->ci_parity_eof = lov_io_parity_eof(obj,
lio->lis_mirror_index, file_size);
CDEBUG(D_INODE, "designated parity IO: mirror_idx=%d, "
"file_size=%llu, parity_eof=%lld\n",
lio->lis_mirror_index, file_size,
io->ci_parity_eof);
}
RETURN(0);
}
result = lov_io_mirror_write_intent(lio, obj, io);
if (result)
RETURN(result);
if (io->ci_need_write_intent) {
CDEBUG(D_VFSTRACE, DFID " need write intent for [%llu, %llu)\n",
PFID(lu_object_fid(lov2lu(obj))),
lio->lis_pos, lio->lis_endpos);
if (cl_io_is_trunc(io)) {
/**
* for truncate, we uses [size, EOF) to judge whether
* a write intent needs to be send, but we need to
* restore the write extent to [0, size], in truncate,
* the byte in the size position is accessed.
*/
io->ci_write_intent.e_start = 0;
io->ci_write_intent.e_end =
io->u.ci_setattr.sa_attr.lvb_size + 1;
}
/* stop cl_io_init() loop */
RETURN(1);
}
if (io->ci_ndelay_tried == 0 || /* first time to try */
/* reset the mirror index if layout has changed */
lio->lis_mirror_layout_gen != obj->lo_lsm->lsm_layout_gen) {
lio->lis_mirror_layout_gen = obj->lo_lsm->lsm_layout_gen;
index = lio->lis_mirror_index = comp->lo_preferred_mirror;
} else {
index = lio->lis_mirror_index;
LASSERT(index >= 0);
/* move mirror index to the next one */
index = (index + 1) % comp->lo_mirror_count;
}
for (i = 0; i < comp->lo_mirror_count; i++) {
struct lu_extent ext = { .e_start = lio->lis_pos,
.e_end = lio->lis_pos + 1 };
struct lov_mirror_entry *lre;
struct lov_layout_entry *lle;
bool found = false;
lre = lov_mirror_entry(obj, (index + i) % comp->lo_mirror_count);
if (!lre->lre_valid)
continue;
if (lre->lre_foreign)
continue;
/* skip parity mirrors for read IOs unless designated */
if (lre->lre_parity &&
(io->ci_type == CIT_READ || io->ci_type == CIT_FAULT) &&
!io->ci_designated_mirror) {
skipped_parity = true;
continue;
}
lov_foreach_mirror_layout_entry(obj, lle, lre) {
if (!lle->lle_valid)
continue;
if (lu_extent_is_overlapped(&ext, lle->lle_extent)) {
found = true;
if (!io->ci_cross_ec &&
lle->lle_lsme->lsme_dstripe_count != 0)
io->ci_cross_ec = 1;
break;
}
} /* each component of the mirror */
if (found) {
index = (index + i) % comp->lo_mirror_count;
break;
}
} /* each mirror */
if (i == comp->lo_mirror_count) {
/* If we only skipped parity mirrors, return EINVAL */
if (skipped_parity) {
CERROR(DFID": only parity mirrors available for read I/O at %llu\n",
PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
RETURN(-EINVAL);
}
CERROR(DFID": failed to find a component covering I/O region at %llu\n",
PFID(lu_object_fid(lov2lu(obj))), lio->lis_pos);
dump_lsm(D_ERROR, obj->lo_lsm);
RETURN(-EIO);
}
CDEBUG(D_VFSTRACE, DFID ": flr state: %d, move mirror from %d to %d, have retried: %d, mirror count: %d\n",
PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj),
lio->lis_mirror_index, index, io->ci_ndelay_tried,
comp->lo_mirror_count);
lio->lis_mirror_index = index;
/* Set ci_parity_io if this is a parity mirror */
io->ci_parity_io = lov_mirror_entry(obj, index)->lre_parity;
/* Calculate parity EOF if this is a parity mirror */
if (io->ci_parity_io) {
loff_t file_size;
/*
* Get file size from the top (VVP) layer only.
* See comment in designated mirror path above.
*/
file_size = lov_io_top_file_size(env, obj);
io->ci_parity_eof = lov_io_parity_eof(obj, index, file_size);
CDEBUG(D_INODE, "mirror selection parity IO: mirror_idx=%d, "
"file_size=%llu, parity_eof=%lld\n",
index, file_size, io->ci_parity_eof);
}
/* we can't use parity mirrors for write unless designated */
if (lov_mirror_entry(obj, index)->lre_parity &&
io->ci_type == CIT_WRITE &&
io->ci_designated_mirror == 0) {
CERROR(DFID": trying to use parity mirror %d for write\n",
PFID(lu_object_fid(lov2lu(obj))), index);
RETURN(-EINVAL);
}
/*
* FLR: if all mirrors have been tried once, most likely the network
* of this client has been partitioned. We should relinquish CPU for
* a while before trying again.
*/
ndelay_retry:
if (io->ci_ndelay && io->ci_ndelay_tried > 0 &&
(io->ci_ndelay_tried % comp->lo_mirror_count == 0)) {
schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
if (signal_pending(current))
RETURN(-EINTR);
/**
* we'd set ci_tried_all_mirrors to turn off fast mirror
* switching for read after we've tried all mirrors several
* rounds.
*/
io->ci_tried_all_mirrors = io->ci_ndelay_tried %
(comp->lo_mirror_count * 4) == 0;
/* if the read crosses possible erasure code, we'd change
* to CIT_EC_RD trying to recover data from parity objects.
*/
if (io->ci_type == CIT_READ && io->ci_cross_ec) {
CDEBUG(D_VFSTRACE, DFID " switch to CIT_EC_RD\n",
PFID(lu_object_fid(lov2lu(obj))));
io->ci_switch_ec_io = 1;
io->ci_need_restart = 1;
io->ci_ndelay_tried = 0;
RETURN(-ENODATA);
}
}
++io->ci_ndelay_tried;
CDEBUG(D_VFSTRACE, "use %sdelayed RPC state for this IO\n",
io->ci_ndelay ? "non-" : "");
RETURN(0);
}
static int lov_io_slice_init(const struct lu_env *env, struct lov_io *lio,
struct lov_object *obj, struct cl_io *io)
{
int index;
int result = 0;
bool rdonly;
ENTRY;
io->ci_result = 0;
lio->lis_object = obj;
lio->lis_cached_entry = LIS_CACHE_ENTRY_NONE;
rdonly = lsm_is_rdonly(obj->lo_lsm);
switch (io->ci_type) {
case CIT_EC_RD:
io->u.ci_ec.ec_inner = io->u.ci_ec.ec_outer;
fallthrough;
case CIT_READ:
case CIT_WRITE:
if (io->ci_type == CIT_WRITE && rdonly) {
io->ci_need_pccro_clear = 1;
GOTO(out, result = 1);
}
lio->lis_pos = io->u.ci_rw.crw_pos;
lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_bytes;
lio->lis_io_endpos = lio->lis_endpos;
if (io->ci_type == CIT_EC_RD)
CDEBUG(D_VFSTRACE, "ec read outer [%llu, %llu-%llu)\n",
lio->lis_pos, lio->lis_endpos,
lio->lis_io_endpos);
if (cl_io_is_append(io)) {
/*
* If there is LOV EA hole, then we may cannot locate
* the current file-tail exactly.
*/
if (unlikely(obj->lo_lsm->lsm_entries[0]->lsme_pattern &
LOV_PATTERN_F_HOLE))
GOTO(out, result = -EIO);
}
break;
case CIT_SETATTR:
if (cl_io_is_fallocate(io)) {
if (rdonly) {
io->ci_need_pccro_clear = 1;
GOTO(out, result = 1);
}
lio->lis_pos = io->u.ci_setattr.sa_falloc_offset;
lio->lis_endpos = io->u.ci_setattr.sa_falloc_end;
} else if (cl_io_is_trunc(io)) {
if (rdonly) {
io->ci_need_pccro_clear = 1;
GOTO(out, result = 1);
}
lio->lis_pos = io->u.ci_setattr.sa_attr.lvb_size;
lio->lis_endpos = OBD_OBJECT_EOF;
} else {
lio->lis_pos = 0;
lio->lis_endpos = OBD_OBJECT_EOF;
}
break;
case CIT_DATA_VERSION:
lio->lis_pos = 0;
lio->lis_endpos = OBD_OBJECT_EOF;
break;
case CIT_FAULT: {
pgoff_t index = io->u.ci_fault.ft_index;
if (cl_io_is_mkwrite(io) && rdonly) {
io->ci_need_pccro_clear = 1;
GOTO(out, result = -ENODATA);
}
lio->lis_pos = index << PAGE_SHIFT;
lio->lis_endpos = (index + 1) << PAGE_SHIFT;
break;
}
case CIT_FSYNC: {
lio->lis_pos = io->u.ci_fsync.fi_start;
lio->lis_endpos = io->u.ci_fsync.fi_end;
break;
}
case CIT_LADVISE: {
lio->lis_pos = io->u.ci_ladvise.lio_start;
lio->lis_endpos = io->u.ci_ladvise.lio_end;
break;
}
case CIT_LSEEK: {
lio->lis_pos = io->u.ci_lseek.ls_start;
lio->lis_endpos = OBD_OBJECT_EOF;
break;
}
case CIT_GLIMPSE:
lio->lis_pos = 0;
lio->lis_endpos = OBD_OBJECT_EOF;
break;
case CIT_MISC:
lio->lis_pos = 0;
lio->lis_endpos = OBD_OBJECT_EOF;
break;
default:
LBUG();
}
/*
* CIT_MISC + ci_ignore_layout can identify the I/O from the OSC layer,
* it won't care/access lov layout related info.
*/
if (io->ci_ignore_layout && io->ci_type == CIT_MISC)
GOTO(out, result = 0);
LASSERT(obj->lo_lsm != NULL);
result = lov_io_mirror_init(env, lio, obj, io);
if (result)
GOTO(out, result);
/* check if it needs to instantiate layout */
if (!(io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io) ||
cl_io_is_fallocate(io) ||
(cl_io_is_trunc(io) && io->u.ci_setattr.sa_attr.lvb_size > 0)))
GOTO(out, result = 0);
/*
* for truncate, it only needs to instantiate the components
* before the truncated size.
*/
if (cl_io_is_trunc(io)) {
io->ci_write_intent.e_start = 0;
/* for writes, e_end is endpos, the location of the file
* pointer after the write is completed, so it is not accessed.
* For truncate, 'end' is the size, and *is* acccessed.
* In other words, writes are [start, end), but truncate is
* [start, size], where both are included. So add 1 to the
* size when creating the write intent to account for this.
*/
io->ci_write_intent.e_end =
io->u.ci_setattr.sa_attr.lvb_size + 1;
} else {
io->ci_write_intent.e_start = lio->lis_pos;
io->ci_write_intent.e_end = lio->lis_endpos;
}
CDEBUG(D_LAYOUT, "%llu %llu\n", io->ci_write_intent.e_start,
io->ci_write_intent.e_end);
index = 0;
lov_foreach_io_layout(index, lio, &io->ci_write_intent) {
if (!lsm_entry_inited(obj->lo_lsm, index)) {
io->ci_need_write_intent = 1;
break;
}
}
if (io->ci_need_write_intent && io->ci_designated_mirror > 0) {
/*
* REINT_SYNC RPC has already tried to instantiate all of the
* components involved, obviously it didn't succeed. Skip this
* mirror for now. The server won't be able to figure out
* which mirror it should instantiate components
*/
CERROR(DFID": trying to instantiate components for designated I/O, file state: %d\n",
PFID(lu_object_fid(lov2lu(obj))), lov_flr_state(obj));
io->ci_need_write_intent = 0;
GOTO(out, result = -EIO);
}
if (io->ci_need_write_intent)
GOTO(out, result = 1);
EXIT;
out:
return result;
}
static void lov_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct lov_object *lov = cl2lov(ios->cis_obj);
struct lov_io_sub *sub;
struct cl_io *io = lio->lis_cl.cis_io;
ENTRY;
LASSERT(list_empty(&lio->lis_active));
while ((sub = list_first_entry_or_null(&lio->lis_subios,
struct lov_io_sub,
sub_list)) != NULL) {
list_del_init(&sub->sub_list);
lio->lis_nr_subios--;
lov_io_sub_fini(env, lio, sub);
lov_sub_free(lio, sub);
}
LASSERT(lio->lis_nr_subios == 0);
if (!(io->ci_ignore_layout && io->ci_type == CIT_MISC)) {
LASSERT(atomic_read(&lov->lo_active_ios) > 0);
if (atomic_dec_and_test(&lov->lo_active_ios))
wake_up(&lov->lo_waitq);
}
EXIT;
}
static void lov_io_sub_inherit(struct lov_io_sub *sub, struct lov_io *lio,
loff_t start, loff_t end)
{
struct cl_io *io = &sub->sub_io;
struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
struct cl_io *parent = lio->lis_cl.cis_io;
int index = lov_comp_entry(sub->sub_subio_index);
int stripe = lov_comp_stripe(sub->sub_subio_index);
switch (io->ci_type) {
case CIT_SETATTR: {
io->u.ci_setattr.sa_attr = parent->u.ci_setattr.sa_attr;
io->u.ci_setattr.sa_attr_flags =
parent->u.ci_setattr.sa_attr_flags;
io->u.ci_setattr.sa_avalid = parent->u.ci_setattr.sa_avalid;
io->u.ci_setattr.sa_xvalid = parent->u.ci_setattr.sa_xvalid;
io->u.ci_setattr.sa_falloc_mode =
parent->u.ci_setattr.sa_falloc_mode;
io->u.ci_setattr.sa_stripe_index = stripe;
io->u.ci_setattr.sa_parent_fid =
parent->u.ci_setattr.sa_parent_fid;
/* For SETATTR(fallocate) pass the subtype to lower IO */
io->u.ci_setattr.sa_subtype = parent->u.ci_setattr.sa_subtype;
if (cl_io_is_fallocate(io)) {
io->u.ci_setattr.sa_falloc_offset = start;
io->u.ci_setattr.sa_falloc_end = end;
io->u.ci_setattr.sa_attr_uid =
parent->u.ci_setattr.sa_attr_uid;
io->u.ci_setattr.sa_attr_gid =
parent->u.ci_setattr.sa_attr_gid;
io->u.ci_setattr.sa_attr_projid =
parent->u.ci_setattr.sa_attr_projid;
}
if (cl_io_is_trunc(io)) {
loff_t new_size = parent->u.ci_setattr.sa_attr.lvb_size;
new_size = lov_size_to_stripe(lsm, index, new_size,
stripe);
io->u.ci_setattr.sa_attr.lvb_size = new_size;
io->u.ci_setattr.sa_attr_uid =
parent->u.ci_setattr.sa_attr_uid;
io->u.ci_setattr.sa_attr_gid =
parent->u.ci_setattr.sa_attr_gid;
io->u.ci_setattr.sa_attr_projid =
parent->u.ci_setattr.sa_attr_projid;
}
lov_lsm2layout(lsm, lsm->lsm_entries[index],
&io->u.ci_setattr.sa_layout);
break;
}
case CIT_DATA_VERSION: {
io->u.ci_data_version.dv_data_version = 0;
io->u.ci_data_version.dv_flags =
parent->u.ci_data_version.dv_flags;
break;
}
case CIT_FAULT: {
loff_t off = parent->u.ci_fault.ft_index << PAGE_SHIFT;
io->u.ci_fault = parent->u.ci_fault;
off = lov_size_to_stripe(lsm, index, off, stripe);
io->u.ci_fault.ft_index = off >> PAGE_SHIFT;
break;
}
case CIT_FSYNC: {
io->u.ci_fsync.fi_start = start;
io->u.ci_fsync.fi_end = end;
io->u.ci_fsync.fi_fid = parent->u.ci_fsync.fi_fid;
io->u.ci_fsync.fi_mode = parent->u.ci_fsync.fi_mode;
io->u.ci_fsync.fi_prio = parent->u.ci_fsync.fi_prio;
break;
}
case CIT_READ:
case CIT_WRITE: {
io->u.ci_wr.wr_sync = cl_io_is_sync_write(parent);
io->ci_tried_all_mirrors = parent->ci_tried_all_mirrors;
if (cl_io_is_append(parent))
io->u.ci_wr.wr_append = 1;
io->u.ci_rw.crw_pos = start;
io->u.ci_rw.crw_bytes = end - start;
break;
}
case CIT_EC_RD: {
io->u.ci_rw.crw_pos = start;
io->u.ci_rw.crw_bytes = end - start;
break;
}
case CIT_LADVISE: {
io->u.ci_ladvise.lio_start = start;
io->u.ci_ladvise.lio_end = end;
io->u.ci_ladvise.lio_fid = parent->u.ci_ladvise.lio_fid;
io->u.ci_ladvise.lio_advice = parent->u.ci_ladvise.lio_advice;
io->u.ci_ladvise.lio_flags = parent->u.ci_ladvise.lio_flags;
break;
}
case CIT_LSEEK: {
io->u.ci_lseek.ls_start = start;
io->u.ci_lseek.ls_whence = parent->u.ci_lseek.ls_whence;
io->u.ci_lseek.ls_result = parent->u.ci_lseek.ls_result;
break;
}
case CIT_GLIMPSE:
case CIT_MISC:
default:
break;
}
}
static loff_t lov_offset_mod(loff_t val, int delta)
{
if (val != OBD_OBJECT_EOF)
val += delta;
return val;
}
static int lov_io_add_sub(const struct lu_env *env, struct lov_io *lio,
struct lov_io_sub *sub, u64 start, u64 end)
{
int rc;
end = lov_offset_mod(end, 1);
lov_io_sub_inherit(sub, lio, start, end);
rc = cl_io_iter_init(sub->sub_env, &sub->sub_io);
if (rc != 0) {
cl_io_iter_fini(sub->sub_env, &sub->sub_io);
return rc;
}
list_add_tail(&sub->sub_linkage, &lio->lis_active);
return rc;
}
static int lov_io_iter_init(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
bool is_trunc = cl_io_is_trunc(ios->cis_io);
struct lov_io_sub *sub;
struct lu_extent ext;
int index;
int rc = 0;
ENTRY;
ext.e_start = lio->lis_pos;
ext.e_end = lio->lis_endpos;
if (is_trunc) {
OBD_ALLOC_PTR_ARRAY(lio->lis_trunc_stripe_index,
lio->lis_object->u.composite.lo_entry_count);
if (lio->lis_trunc_stripe_index == NULL)
RETURN(-ENOMEM);
}
lov_foreach_io_layout(index, lio, &ext) {
struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
struct lov_layout_raid0 *r0 = &le->lle_raid0;
u64 start;
u64 end;
int stripe;
bool tested_trunc_stripe = false;
if (is_trunc)
lio->lis_trunc_stripe_index[index] = -1;
CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
index, lsm->lsm_entries[index]->lsme_flags);
if (!lsm_entry_inited(lsm, index)) {
/*
* Read from uninitialized components should return
* zero filled pages.
*/
continue;
}
if (lsm_entry_is_foreign(lsm, index))
continue;
if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
CERROR("I/O to invalid component: %d, mirror: %d\n",
index, lio->lis_mirror_index);
RETURN(-EIO);
}
for (stripe = 0; stripe < r0->lo_nr; stripe++) {
if (!lov_stripe_intersects(lsm, index, stripe,
&ext, &start, &end))
continue;
if (unlikely(!r0->lo_sub[stripe])) {
if (ios->cis_io->ci_type == CIT_READ ||
ios->cis_io->ci_type == CIT_WRITE ||
ios->cis_io->ci_type == CIT_FAULT)
RETURN(-EIO);
continue;
}
if (is_trunc && !tested_trunc_stripe) {
int prev;
u64 tr_start;
prev = (stripe == 0) ? r0->lo_nr - 1 :
stripe - 1;
/**
* Only involving previous stripe if the
* truncate in this component is at the
* beginning of this stripe.
*/
tested_trunc_stripe = true;
if (ext.e_start < lsm->lsm_entries[index]->
lsme_extent.e_start) {
/* need previous stripe involvement */
lio->lis_trunc_stripe_index[index] = prev;
} else {
div64_u64_rem(ext.e_start,
stripe_width(lsm, index),
&tr_start);
/* tr_start %= stripe_swidth */
if (tr_start == stripe * lsm->
lsm_entries[index]->
lsme_stripe_size)
lio->lis_trunc_stripe_index[index] = prev;
}
}
/* if the last stripe is the trunc stripeno */
if (is_trunc &&
lio->lis_trunc_stripe_index[index] == stripe)
lio->lis_trunc_stripe_index[index] = -1;
sub = lov_sub_get(env, lio,
lov_comp_index(index, stripe));
if (IS_ERR(sub))
return PTR_ERR(sub);
rc = lov_io_add_sub(env, lio, sub, start, end);
if (rc != 0)
break;
}
if (rc != 0)
break;
if (is_trunc && lio->lis_trunc_stripe_index[index] != -1) {
stripe = lio->lis_trunc_stripe_index[index];
if (unlikely(!r0->lo_sub[stripe])) {
lio->lis_trunc_stripe_index[index] = -1;
continue;
}
sub = lov_sub_get(env, lio,
lov_comp_index(index, stripe));
if (IS_ERR(sub))
return PTR_ERR(sub);
/**
* the prev sub could be used by another truncate, we'd
* skip it. LU-14128 happends when expand truncate +
* read get wrong kms.
*/
if (!list_empty(&sub->sub_linkage)) {
lio->lis_trunc_stripe_index[index] = -1;
continue;
}
(void)lov_stripe_intersects(lsm, index, stripe, &ext,
&start, &end);
rc = lov_io_add_sub(env, lio, sub, start, end);
if (rc != 0)
break;
}
}
RETURN(rc);
}
static int lov_io_rw_iter_init(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *io = ios->cis_io;
struct lov_stripe_md_entry *lse;
loff_t start = io->u.ci_rw.crw_pos;
loff_t next;
int index;
LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
ENTRY;
if (cl_io_is_append(io)) {
struct lov_layout_entry *lle;
struct lov_mirror_entry *lre;
loff_t endpos = 0;
/* cover the whole inited region for append */
LASSERT(lio->lis_mirror_index >= 0);
lre = lov_mirror_entry(lio->lis_object, lio->lis_mirror_index);
lov_foreach_mirror_layout_entry(lio->lis_object, lle, lre) {
if (!lsme_inited(lle->lle_lsme))
break;
endpos = lle->lle_lsme->lsme_extent.e_end;
}
lio->lis_pos = 0;
if (endpos == OBD_OBJECT_EOF || lio->lis_endpos < endpos)
lio->lis_endpos = endpos;
io->u.ci_wr.wr_append_lockpos = lio->lis_endpos;
RETURN(lov_io_iter_init(env, ios));
}
index = lov_io_layout_at(lio, io->u.ci_rw.crw_pos);
if (index < 0) { /* non-existing layout component */
if (io->ci_type == CIT_READ) {
/*
* TODO: it needs to detect the next component and
* then set the next pos
*/
io->ci_continue = 0;
RETURN(lov_io_iter_init(env, ios));
}
RETURN(-ENODATA);
}
if (!lov_entry(lio->lis_object, index)->lle_valid &&
!io->ci_designated_mirror)
RETURN(io->ci_type == CIT_READ ? -EAGAIN : -EIO);
lse = lov_lse(lio->lis_object, index);
if (lsme_is_foreign(lse))
RETURN(-EINVAL);
next = MAX_LFS_FILESIZE;
if (lse->lsme_stripe_count > 1) {
unsigned long ssize = lse->lsme_stripe_size;
start = div64_u64(start, ssize);
next = (start + 1) * ssize;
if (next <= start * ssize)
next = MAX_LFS_FILESIZE;
}
LASSERTF(io->u.ci_rw.crw_pos >= lse->lsme_extent.e_start,
"pos %lld, [%lld, %lld)\n", io->u.ci_rw.crw_pos,
lse->lsme_extent.e_start, lse->lsme_extent.e_end);
next = min_t(__u64, next, lse->lsme_extent.e_end);
next = min_t(__u64, next, lio->lis_io_endpos);
io->ci_continue = next < lio->lis_io_endpos;
io->u.ci_rw.crw_bytes = next - io->u.ci_rw.crw_pos;
lio->lis_pos = io->u.ci_rw.crw_pos;
lio->lis_endpos = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_bytes;
CDEBUG(D_VFSTRACE,
"stripe: %llu chunk: [%llu, %llu) %llu, %zd\n",
(__u64)start, lio->lis_pos, lio->lis_endpos,
(__u64)lio->lis_io_endpos, io->u.ci_rw.crw_bytes);
/*
* XXX The following call should be optimized: we know, that
* [lio->lis_pos, lio->lis_endpos) intersects with exactly one stripe.
*/
RETURN(lov_io_iter_init(env, ios));
}
static int lov_ecio_add_data_sub(const struct lu_env *env, struct lov_io *lio,
struct lu_extent *ext, int *err_nr)
{
struct cl_io *io = lio->lis_cl.cis_io;
struct lov_object *lov = lio->lis_object;
struct lov_device *dev = lov_object_dev(lov);
struct lov_stripe_md *lsm = lov->lo_lsm;
struct lov_layout_entry *lle;
struct lov_layout_raid0 *r0;
struct lu_extent data_ext;
loff_t file_size;
int index;
int rc = 0;
ENTRY;
/*
* Cap the extent to the actual file size. lov_io_set_range() expands
* the extent to cover full RAID sets for parity calculation, but data
* stripes beyond EOF contain no data. If such a stripe is on an
* unavailable OST it would be falsely marked degraded, triggering
* recovery that corrupts valid data.
*/
file_size = io->u.ci_ec.ec_inode_size;
data_ext = *ext;
if (data_ext.e_end > file_size)
data_ext.e_end = file_size;
if (data_ext.e_end <= data_ext.e_start)
RETURN(0);
lov_foreach_io_layout(index, lio, &data_ext) {
unsigned int stripe;
if (!lsm_entry_inited(lsm, index))
continue;
lle = lov_entry(lov, index);
if (!lle->lle_valid) {
CWARN("%s: "DFID": I/O on invalid component %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))), index, -EIO);
RETURN(-EIO);
}
r0 = &lle->lle_raid0;
for (stripe = 0; stripe < r0->lo_nr; stripe++) {
struct lov_io_sub *sub;
u64 sub_start;
u64 sub_end;
if (!lov_stripe_intersects(lsm, index, stripe,
&data_ext,
&sub_start, &sub_end)) {
/* Clear any READ_ERR set by the lock path
* for stripes not covered by this loop - no
* data means no recovery needed.
*/
if (r0->lo_sub[stripe])
r0->lo_sub[stripe]->lso_status =
LSS_OK;
continue;
}
if (unlikely(!r0->lo_sub[stripe])) {
CWARN("%s: "DFID": stripe %u in comp %d is NULL, rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))),
stripe, index, -EIO);
RETURN(-EIO);
}
sub = lov_sub_get(env, lio,
lov_comp_index(index, stripe));
if (IS_ERR(sub)) {
r0->lo_sub[stripe]->lso_status = LSS_READ_ERR;
(*err_nr)++;
CDEBUG(D_INFO, DFID": stripe %u in comp %d is unavailable (%d): rc = %ld\n",
PFID(lu_object_fid(lov2lu(lov))),
stripe, index, *err_nr, PTR_ERR(sub));
continue;
}
rc = lov_io_add_sub(env, lio, sub, sub_start, sub_end);
if (rc != 0) {
r0->lo_sub[stripe]->lso_status = LSS_READ_ERR;
(*err_nr)++;
CDEBUG(D_INFO, DFID": stripe %u in comp %d is unavailable (%d): rc = %d\n",
PFID(lu_object_fid(lov2lu(lov))),
stripe, index, *err_nr, rc);
rc = 0;
continue;
} else {
r0->lo_sub[stripe]->lso_status = LSS_OK;
}
}
}
RETURN(rc);
}
static int lov_ecio_add_parity_sub(const struct lu_env *env, struct lov_io *lio,
struct lu_extent *ext,
struct ec_split_comp *sc, int nr)
{
struct lov_object *lov = lio->lis_object;
struct lov_device *dev = lov_object_dev(lov);
struct lov_stripe_md *lsm = lov->lo_lsm;
struct lov_layout_entry *lle;
struct lov_layout_raid0 *r0;
int parity_mir_idx;
int index;
int rc = 0;
ENTRY;
if (!lov->u.composite.lo_entries ||
lov->u.composite.lo_entry_count == 0)
RETURN(rc);
parity_mir_idx = lov_parity_mirror_index_from_data(lio,
lio->lis_mirror_index);
if (parity_mir_idx < 0) {
CWARN("%s: "DFID": failed to find valid parity mirror for data mirror %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))), lio->lis_mirror_index,
parity_mir_idx);
RETURN(-EINVAL);
}
lov_foreach_io_layout_mirror(index, lio, ext, parity_mir_idx) {
struct lov_stripe_md_entry *lsme;
struct lov_io_sub *sub;
u64 offset;
u64 base_poff;
u64 dswidth;
u64 sub_start;
u64 sub_end;
u16 dcount;
int stripe;
int added = 0;
lle = lov_entry(lov, index);
r0 = &lle->lle_raid0;
if (!lsm_entry_inited(lsm, index))
continue;
if (!lle->lle_valid ||
!(lle->lle_lsme->lsme_flags & LCME_FL_PARITY)) {
CWARN("%s: "DFID": I/O on invalid parity component %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))), index, -EIO);
RETURN(-EIO);
}
if (nr > r0->lo_nr) {
CWARN("%s: "DFID": too few parity (%d) components to recover data (%d): rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))), r0->lo_nr, nr,
-EAGAIN);
RETURN(-EAGAIN);
}
lsme = lle->lle_lsme;
/*
* Parity sub-objects use lov_stripe_offset() addressing:
* file offset e_start maps to sub-object offset
* (e_start / parity_swidth) * stripe_size.
* Compute this base for PFL components that don't start at
* file offset 0.
*/
base_poff = div64_u64(lle->lle_extent->e_start,
(u64)lsme->lsme_stripe_count *
lsme->lsme_stripe_size) *
lsme->lsme_stripe_size;
/*
* Map data file offsets to parity sub-object offsets.
* Each data RAID row (data_stripe_count * stripe_size
* bytes) produces exactly one stripe_size of parity
* per parity sub-object, so:
* sub_obj_offset = data_row * stripe_size
*
* The original code divided by parity swidth instead
* of data swidth, producing wildly wrong offsets when
* data_stripe_count != parity_stripe_count.
*/
dcount = sc->esc_k0 * sc->esc_n0 +
sc->esc_k1 * sc->esc_n1;
dswidth = (u64)dcount * lsme->lsme_stripe_size;
/* Defense-in-depth: a non-EC cycle reaches here with a zeroed
* @sc (all esc_* == 0), giving dcount == 0 and dswidth == 0.
* lov_io_ec_rd_iter_init() now skips this function for
* dstripe_count == 0, but guard the div64_u64() below so a
* bad caller cannot trap on a divide by zero.
*/
if (dswidth == 0)
continue;
/* sub_start: floor(data_offset / data_swidth) rows */
offset = ext->e_start - lle->lle_extent->e_start;
sub_start = base_poff +
div64_u64(offset, dswidth) *
lsme->lsme_stripe_size;
/* sub_end: ceil(data_offset / data_swidth) rows */
if (ext->e_end > lle->lle_extent->e_end)
offset = lle->lle_extent->e_end -
lle->lle_extent->e_start;
else
offset = ext->e_end - lle->lle_extent->e_start;
sub_end = base_poff +
div64_u64(offset + dswidth - 1, dswidth) *
lsme->lsme_stripe_size;
for (stripe = 0; stripe < r0->lo_nr; stripe++) {
if (unlikely(!r0->lo_sub[stripe]))
continue;
sub = lov_sub_get(env, lio,
lov_comp_index(index, stripe));
if (IS_ERR(sub))
continue;
/* page read from parity component doesn't cover with
* ldlm lock
*/
sub->sub_io.ci_lockless = 1;
rc = lov_io_add_sub(env, lio, sub, sub_start, sub_end);
if (rc != 0) {
rc = 0;
continue;
}
added++;
}
if (added < nr) {
CWARN("%s: "DFID": available parity stripes %d < needed %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))),
added, nr, -EIO);
RETURN(-EIO);
}
}
RETURN(rc);
}
static int lov_io_ec_rd_iter_init(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *io = ios->cis_io;
struct lov_object *lov = lio->lis_object;
loff_t start = io->u.ci_ec.ec_inner.crw_pos;
loff_t next; /* file offset of next read cycle */
struct lov_stripe_md_entry *lsme;
struct lu_extent ext;
struct ec_split_comp sc = { 0 };
int index;
int err_nr = 0;
int rc = 0;
ENTRY;
index = lov_io_layout_at(lio, start);
if (index < 0)
RETURN(-ENODATA);
/* use io inner range */
lsme = lov_lse(lov, index);
next = MAX_LFS_FILESIZE;
if (lsme->lsme_dstripe_count >= 1) {
u64 offset;
u64 offset_i;
u64 row;
u64 cnt;
/* set next read offset to next raid set with different EC
* parameter
*/
ec_split_stripes(lsme->lsme_stripe_count,
lsme->lsme_dstripe_count, &sc);
offset = start - lsme->lsme_extent.e_start;
row = div64_u64(offset, (u64)lsme->lsme_stripe_size *
lsme->lsme_stripe_count);
offset_i = div64_u64(offset, lsme->lsme_stripe_size) %
lsme->lsme_stripe_count;
/* cnt is the stripe number from the component start to the
* end of the raid set where @offset locates.
*/
if (offset_i < sc.esc_k0 * sc.esc_n0) {
cnt = row * lsme->lsme_stripe_count +
sc.esc_k0 * sc.esc_n0;
} else {
cnt = row * lsme->lsme_stripe_count +
sc.esc_k0 * sc.esc_n0 + sc.esc_k1 * sc.esc_n1;
}
next = lsme->lsme_extent.e_start + cnt * lsme->lsme_stripe_size;
CDEBUG(D_INFO, DFID
" next sc(%d/%d/%d/%d), row %lld, offset_i %lld, cnt %lld, next %llu\n",
PFID(lu_object_fid(lov2lu(lov))),
sc.esc_k0, sc.esc_n0, sc.esc_k1, sc.esc_n1,
row, offset_i, cnt, next);
}
next = min_t(__u64, next, lsme->lsme_extent.e_end);
next = min_t(loff_t, next, lio->lis_io_endpos);
/* set this read cycle range */
io->ci_continue = next < lio->lis_io_endpos;
io->u.ci_ec.ec_inner.crw_bytes = next - start;
lio->lis_pos = start;
lio->lis_endpos = next;
CDEBUG(D_VFSTRACE, "ec read [%llu, %llu-%llu)\n",
(__u64)start, (__u64)next, (__u64)lio->lis_io_endpos);
ext.e_start = start;
ext.e_end = next;
/* add data component sub io */
rc = lov_ecio_add_data_sub(env, lio, &ext, &err_nr);
if (rc < 0)
RETURN(rc);
/* Pre-add parity sub-IOs for an EC component. Parity reads may be
* triggered later in lov_io_ec_rd_start() by transient BRW read
* errors on live data OSTs (network blips, fault injection), not
* only by imports that were already deactivated at iter_init
* time. lov_sub_get() allocating a fresh parity sub on-demand
* there would miss ci_lockless=1 and skip cl_io_iter_init(), so
* osc_req_attr_set() would LBUG "uncovered page!" because parity
* OSCs have no DLM lock enqueued.
*/
if (lsme->lsme_dstripe_count >= 1) {
rc = lov_ecio_add_parity_sub(env, lio, &ext, &sc, err_nr);
if (rc < 0)
RETURN(rc);
}
RETURN(rc);
}
static int lov_io_setattr_iter_init(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *io = ios->cis_io;
int index;
ENTRY;
if (cl_io_is_trunc(io) && lio->lis_pos > 0) {
index = lov_io_layout_at(lio, lio->lis_pos - 1);
/* no entry found for such offset */
if (index < 0)
RETURN(io->ci_result = -ENODATA);
}
RETURN(lov_io_iter_init(env, ios));
}
static int lov_io_call(const struct lu_env *env, struct lov_io *lio,
int (*iofunc)(const struct lu_env *, struct cl_io *))
{
struct cl_io *parent = lio->lis_cl.cis_io;
struct lov_io_sub *sub;
int rc = 0;
ENTRY;
list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
rc = iofunc(sub->sub_env, &sub->sub_io);
if (rc) {
/**
* fsync race with truncate, we'd continue to other
* OST object's fsync to potentially discard
* caching pages (osc_cache_writeback_range).
*/
if (rc == -ENOENT && parent->ci_type == CIT_FSYNC)
continue;
break;
}
if (parent->ci_result == 0)
parent->ci_result = sub->sub_io.ci_result;
}
RETURN(rc);
}
static int lov_io_lock(const struct lu_env *env, const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *io = lio->lis_cl.cis_io;
struct cl_object *obj = ios->cis_obj;
struct cl_lock_descr descr = { 0 };
int rc;
ENTRY;
if (io->ci_type == CIT_EC_RD) {
__u32 enqflags = CEF_MUST | CEF_HEED_ERROR;
descr.cld_obj = obj;
descr.cld_start = io->u.ci_ec.ec_inner.crw_pos >> PAGE_SHIFT;
descr.cld_end = (io->u.ci_ec.ec_inner.crw_pos +
io->u.ci_ec.ec_inner.crw_bytes - 1) >>
PAGE_SHIFT;
descr.cld_mode = CLM_READ;
descr.cld_enq_flags = enqflags;
/* Apply group-lock handling just like vvp_io_one_lock_index().
* If the file is group-locked (indicated by non-zero
* ci_group_gid), we must request a CLM_GROUP lock with the
* correct gid to match the existing group lock, rather than
* a regular LCK_PR which would deadlock against it.
*/
if (io->ci_group_gid != 0) {
descr.cld_mode = CLM_GROUP;
descr.cld_gid = io->ci_group_gid;
enqflags |= CEF_LOCK_MATCH;
descr.cld_enq_flags = enqflags;
}
CDEBUG(D_INFO, DFID": ec read lock "DDESCR"\n",
PFID(lu_object_fid(lov2lu(lio->lis_object))),
PDESCR(&descr));
/* add data stripe locks */
rc = cl_io_lock_alloc_add(env, io, &descr);
if (rc < 0)
RETURN(rc);
}
rc = lov_io_call(env, cl2lov_io(env, ios), cl_io_lock);
RETURN(rc);
}
static int lov_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
{
ENTRY;
RETURN(lov_io_call(env, cl2lov_io(env, ios), cl_io_start));
}
static int lov_io_end_wrapper(const struct lu_env *env, struct cl_io *io)
{
ENTRY;
/*
* It's possible that lov_io_start() wasn't called against this
* sub-io, either because previous sub-io failed, or upper layer
* completed IO.
*/
if (io->ci_state == CIS_IO_GOING)
cl_io_end(env, io);
else
io->ci_state = CIS_IO_FINISHED;
RETURN(0);
}
static int lov_io_iter_fini_wrapper(const struct lu_env *env, struct cl_io *io)
{
cl_io_iter_fini(env, io);
RETURN(0);
}
static int lov_io_unlock_wrapper(const struct lu_env *env, struct cl_io *io)
{
cl_io_unlock(env, io);
RETURN(0);
}
static void lov_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
{
int rc;
/* Before ending each i/o, we must set lis_cached_entry to tell the
* next i/o not to use stale cached lis information.
*/
cl2lov_io(env, ios)->lis_cached_entry = LIS_CACHE_ENTRY_NONE;
rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_end_wrapper);
LASSERT(rc == 0);
}
enum ec_data_pg_state {
EC_DPG_ZERO, /* beyond EOF/RG */
EC_DPG_UPTODATE, /* already in page cache */
EC_DPG_READ, /* read successfully */
EC_DPG_ERROR, /* stripe dead or read failed */
};
/**
* lov_ec_read_stripe_pages() - read all pages for one data stripe across all
* page positions in a recovery group.
*
* Submits all pages for one stripe in a single
* cl_io_submit_sync call instead of one per page.
*
* @env: lu_env context
* @lio: LOV I/O structure
* @comp_i: composite layout component index
* @stripe_i: which data stripe (0..dcount-1)
* @lsme: stripe metadata entry
* @soff: start offset of recovery group
* @eoff: end offset of recovery group
* @pg_start: first page position in stripe
* @npages: number of page positions
* @out_pages: array[npages] of cl_page pointers (output)
* @out_states: array[npages] of page states (output)
* @dlist: cl_page_list for pages that need reading (output)
*
* Returns 0 on success. Pages in state of EC_DPG_UPTODATE are assumed (owned)
* in out_pages[], pages in state of EC_DPG_READ are owned in dlist, and pages
* in state of EC_DPG_ERROR are owned in ec_page_list. Vmpages can be accessed
* via out_pages[]->cp_vmpage.
*/
static int
lov_ec_read_stripe_pages(const struct lu_env *env, struct lov_io *lio,
int comp_i, int stripe_i,
struct lov_stripe_md_entry *lsme,
unsigned long long soff, unsigned long long eoff,
int pg_start, int npages, struct cl_page **out_pages,
enum ec_data_pg_state *out_states,
struct cl_page_list *dlist)
{
struct lov_object *lov = lio->lis_object;
struct cl_object *obj = lov2cl(lov);
struct cl_io *io = lio->lis_cl.cis_io;
struct inode *inode = cl2lov(obj)->lo_inode;
struct lov_layout_raid0 *r0 = lov_r0(lov, comp_i);
unsigned long ss = lsme->lsme_stripe_size;
int base_stripe;
int stripe; /* physical stripe index */
struct lov_io_sub *sub = NULL;
struct cl_2queue *queue;
int need_read = 0;
int j;
int rc = 0;
ENTRY;
base_stripe = lov_stripe_number(lov->lo_lsm, comp_i, soff);
/* stripe_i is the i-th stripe within the recovery group */
stripe = (base_stripe + stripe_i) % r0->lo_nr;
/* Phase 1: grab all pages, classify them */
for (j = 0; j < npages; j++) {
pgoff_t page_idx;
struct cl_page *clpage;
struct page *vmpage;
struct folio *_folio;
out_pages[j] = NULL;
page_idx = ((soff + stripe_i * ss) >> PAGE_SHIFT) +
pg_start + j;
if ((page_idx << PAGE_SHIFT) >= eoff ||
(page_idx << PAGE_SHIFT) >= io->u.ci_ec.ec_inode_size) {
out_states[j] = EC_DPG_ZERO;
continue;
}
_folio = get_folio_grab(inode->i_mapping, page_idx,
FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
mapping_gfp_mask(inode->i_mapping));
if (IS_ERR_OR_NULL(_folio))
GOTO(out_err, rc = -ENOMEM);
vmpage = fpgptr(_folio);
if (vmpage->mapping != inode->i_mapping) {
unlock_page(vmpage);
put_page(vmpage);
GOTO(out_err, rc = -EAGAIN);
}
/* Create cl_page for this page. Determine state based on:
* - PageUptodate: already in cache (UPTODATE)
* - LSS_READ_ERR: stripe is down (ERROR)
* - Otherwise: needs I/O (READ)
*/
clpage = cl_page_find(env, cl_object_top(obj), page_idx, vmpage,
CPT_CACHEABLE);
if (IS_ERR(clpage)) {
unlock_page(vmpage);
put_page(vmpage);
GOTO(out_err, rc = PTR_ERR(clpage));
}
cl_page_assume(env, io, clpage);
out_pages[j] = clpage;
/* drop vmpage refcount added by get_folio_grab() */
folio_put(_folio);
if (PageUptodate(vmpage)) {
out_states[j] = EC_DPG_UPTODATE;
} else if (r0->lo_sub[stripe]->lso_status == LSS_READ_ERR) {
out_states[j] = EC_DPG_ERROR;
cl_page_list_add(&io->u.ci_ec.ec_page_list, clpage,
false);
} else {
out_states[j] = EC_DPG_READ;
need_read++;
}
}
if (need_read == 0)
RETURN(0);
/* Phase 2: submit all READ pages in one batch */
sub = lov_sub_get(env, lio, lov_comp_index(comp_i, stripe));
if (IS_ERR(sub)) {
/* Could not get a sub-io for this stripe. The cl_pages in
* out_pages[j] are assumed (owned) but never submitted, so
* move them to ec_page_list (matching the submit-failure
* path).
*/
for (j = 0; j < npages; j++) {
if (out_states[j] == EC_DPG_READ) {
cl_page_list_add(&io->u.ci_ec.ec_page_list,
out_pages[j], false);
out_states[j] = EC_DPG_ERROR;
}
}
RETURN(0);
}
queue = &lov_env_info(sub->sub_env)->lti_cl2q;
cl_2queue_init(queue);
for (j = 0; j < npages; j++) {
if (out_states[j] == EC_DPG_READ)
cl_page_list_add(&queue->c2_qin,
out_pages[j], false);
}
rc = cl_io_submit_sync(sub->sub_env, &sub->sub_io,
CRT_READ, queue, 0);
if (rc) {
r0->lo_sub[stripe]->lso_status = LSS_READ_ERR;
/* EC_DPG_READ cl_pages read error, move them to ec_page_list */
cl_page_list_splice(&queue->c2_qin, &io->u.ci_ec.ec_page_list);
cl_page_list_splice(&queue->c2_qout, &io->u.ci_ec.ec_page_list);
cl_2queue_fini(env, queue);
for (j = 0; j < npages; j++) {
if (out_states[j] == EC_DPG_READ)
out_states[j] = EC_DPG_ERROR;
}
RETURN(0); /* not fatal, can try parity */
}
/* Success: pages are in c2_qout, assumed. Splice to dlist and NULL
* out_pages
*/
for (j = 0; j < npages; j++) {
if (out_states[j] == EC_DPG_READ)
SetPageUptodate(out_pages[j]->cp_vmpage);
}
cl_page_list_splice(&queue->c2_qin, dlist);
cl_page_list_splice(&queue->c2_qout, dlist);
cl_2queue_fini(env, queue);
RETURN(0);
out_err:
/* Release pages grabbed before the error.
* - UPTODATE: valid in the page cache, disown (unlock vmpage) and
* drop the cl_page ref. No discard.
* - READ: assumed/owned but never submitted, so not in dlist nor
* ec_page_list. Disown to unlock the vmpage and drop the cl_page
* ref so neither leaks.
* - ERROR: already in ec_page_list; vvp_io_ec_rd_end will disown and
* fini it, so just NULL the pointer.
*/
for (j = 0; j < npages; j++) {
if (out_pages[j]) {
if (out_states[j] == EC_DPG_UPTODATE ||
out_states[j] == EC_DPG_READ) {
cl_page_disown(env, io, out_pages[j]);
cl_page_put(env, out_pages[j]);
}
out_pages[j] = NULL;
}
}
RETURN(rc);
}
/**
* lov_ec_read_parity_stripe() - read all parity pages for one parity stripe
* across all page positions in a recovery group.
*
* @parity_idx: which parity stripe (0..pcount-1)
* @pg_start: first page position
* @npages: number of page positions
* @out_vmpages: array[npages] of vmpage pointers (output)
* @out_clpages: array[npages] of cl_page pointers (output)
* @plist: page list to track allocated pages for cleanup
*/
static int
lov_ec_read_parity_stripe(const struct lu_env *env, struct lov_io *lio,
int pcomp_i, int parity_idx,
struct lov_stripe_md_entry *lsme,
unsigned long long parity_foff, int pg_start,
int npages, struct page **out_vmpages,
struct cl_page **out_clpages,
struct cl_page_list *plist)
{
struct lov_object *lov = lio->lis_object;
struct cl_object *obj = lov2cl(lov);
struct lov_layout_raid0 *r0 = lov_r0(lov, pcomp_i);
int start_stripe;
int stripe;
struct lov_io_sub *sub;
struct cl_object *subobj;
struct cl_2queue *queue;
struct cl_sync_io anchor;
u64 p_foff;
loff_t obdoff;
int j;
int rc = 0;
ENTRY;
p_foff = parity_foff + (u64)parity_idx * lsme->lsme_stripe_size;
start_stripe = lov_stripe_number(lov->lo_lsm, pcomp_i, parity_foff);
stripe = (start_stripe + parity_idx) % r0->lo_nr;
lov_stripe_offset(lov->lo_lsm, pcomp_i, p_foff, stripe, &obdoff);
if (unlikely(!r0->lo_sub[stripe]))
RETURN(-EIO);
sub = lov_sub_get(env, lio, lov_comp_index(pcomp_i, stripe));
if (IS_ERR(sub))
RETURN(-EIO);
subobj = lovsub2cl(r0->lo_sub[stripe]);
/* Phase 1: allocate all pages */
for (j = 0; j < npages; j++) {
pgoff_t page_idx = (obdoff >> PAGE_SHIFT) + pg_start + j;
struct page *vmpage;
vmpage = alloc_page(GFP_NOFS);
if (!vmpage)
GOTO(out_err, rc = -ENOMEM);
lock_page(vmpage);
out_vmpages[j] = vmpage;
out_clpages[j] = cl_page_alloc_sub(env, sub->sub_env, obj,
subobj, page_idx, vmpage,
CPT_TRANSIENT);
if (IS_ERR(out_clpages[j])) {
rc = PTR_ERR(out_clpages[j]);
out_clpages[j] = NULL;
unlock_page(vmpage);
__free_page(vmpage);
out_vmpages[j] = NULL;
GOTO(out_err, rc);
}
}
/* Phase 2: submit all pages in one batch.
* Parity pages are CPT_TRANSIENT, so we must use
* cl_io_submit_rw + cl_sync_io_wait rather than
* cl_io_submit_sync (which asserts !CPT_TRANSIENT).
*/
queue = &lov_env_info(sub->sub_env)->lti_cl2q;
cl_2queue_init(queue);
for (j = 0; j < npages; j++) {
out_clpages[j]->cp_sync_io = &anchor;
cl_page_list_add(&queue->c2_qin, out_clpages[j], false);
}
cl_sync_io_init(&anchor, npages);
rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io, CRT_READ, queue);
if (rc == 0) {
struct cl_page *clpage;
/* If some pages weren't sent for any reason (e.g.,
* page already up-to-date), count them as completed to
* avoid infinite wait.
*/
cl_page_list_for_each(clpage, &queue->c2_qin)
cl_sync_io_note(env, &anchor, 1);
rc = cl_sync_io_wait(sub->sub_env, &anchor, 0);
}
if (rc) {
/* cl_2queue_fini() puts each cl_page in qin/qout (ref drops
* to 0, cl_page freed). Null out_clpages[] before the
* out_err loop so it doesn't cl_page_put() the freed pages
* again -- that's a refcount underflow + UAF (LU-12668).
* Vmpages are still ours to unlock_page + __free_page.
*/
cl_2queue_fini(env, queue);
for (j = 0; j < npages; j++)
out_clpages[j] = NULL;
GOTO(out_err, rc);
}
/* Success: move pages to plist for tracking */
for (j = 0; j < npages; j++)
SetPageUptodate(out_vmpages[j]);
cl_page_list_splice(&queue->c2_qin, plist);
cl_page_list_splice(&queue->c2_qout, plist);
cl_2queue_fini(env, queue);
RETURN(0);
out_err:
for (j = 0; j < npages; j++) {
if (out_clpages[j]) {
cl_page_put(env, out_clpages[j]);
out_clpages[j] = NULL;
}
if (out_vmpages[j]) {
unlock_page(out_vmpages[j]);
__free_page(out_vmpages[j]);
out_vmpages[j] = NULL;
}
}
RETURN(rc);
}
/**
* struct ec_recover_ctx - precomputed EC recovery tables
*
* The GF decode tables depend only on which stripes are missing and which
* parity stripes are used, not on page data. Precompute once per recovery
* group and reuse for every page position.
*/
struct ec_recover_ctx {
u8 *erc_tbls;
unsigned char erc_dcount;
unsigned char erc_pcount;
unsigned char erc_unavail_nr;
unsigned int erc_tbls_size;
};
/**
* struct lov_ec_rd_state - heap-allocated working buffers for one EC read
* cycle (lov_io_ec_rd_start call). Grouped so alloc/free can be split into
* helpers and later swapped for pre-allocated pools without churning the
* call sites.
*
* @ecr_err_array: indices of unavailable data stripes [unavail_nr]
* @ecr_parity_used: indices of parity stripes used for recovery [unavail_nr]
* @ecr_ptrs: kmapped data page pointers (available + parity) [dcount]
* @ecr_recov_ptrs: kmapped recovery destination pointers [pcount]
* @ecr_zerobuf: PAGE_SIZE zero buffer for EC_DPG_ZERO slots
* @ecr_encode_matrix: Cauchy matrix [dcount * (dcount + pcount)]
* @ecr_dcount: data stripe count for this raid set
* @ecr_pcount: parity stripe count for this raid set
* @ecr_temp_matrix: scratch temp matrix for EC recovery
* [dcount * dcount]
* @ecr_invert_matrix: scratch inverted matrix for EC recovery
* [dcount * dcount]
* @ecr_decode_matrix: scratch decode matrix for EC recovery
* [pcount * dcount]
* @ecr_gf_tbls: GF tables for the RG-wide recovery ctx
* [dcount * pcount * 32]
* @ecr_pp_gf_tbls: GF tables for the per-page recovery ctx
* [dcount * pcount * 32]
*
* @ecr_temp_matrix, @ecr_invert_matrix and @ecr_decode_matrix are pure
* scratch space used only within a single lov_ec_recover_init() call to
* derive the GF tables; they are not read afterwards, so both the
* RG-wide and per-page contexts can safely share one set (allocated
* on-demand on first use). @ecr_gf_tbls and @ecr_pp_gf_tbls, however,
* are kept in independent buffers because @ecr_gf_tbls backs the
* RG-wide ctx (recover_ctx), which must remain valid and unmodified
* across the whole recovery group for fast-path pages, while
* @ecr_pp_gf_tbls backs the per-page ctx that is rebuilt for every
* slow-path page; sharing a single GF-tables buffer would let a
* per-page rebuild silently corrupt the RG-wide tables. Both are
* allocated together on the RG-wide lov_ec_recover_init() call, before
* the kmap loop starts, even though @ecr_pp_gf_tbls is only used by
* later per-page calls inside the loop -- per-page calls happen while
* kmap_local_page() mappings may be atomic (kernels < 5.10), where
* OBD_ALLOC() must not sleep. @ecr_decode_matrix is sized for the
* worst case (pcount) since either context may need an unavail_nr up
* to pcount.
*/
struct lov_ec_rd_state {
int *ecr_err_array;
int *ecr_parity_used;
unsigned char **ecr_ptrs;
unsigned char **ecr_recov_ptrs;
unsigned char *ecr_zerobuf;
u8 *ecr_encode_matrix;
unsigned int ecr_dcount;
unsigned int ecr_pcount;
u8 *ecr_temp_matrix;
u8 *ecr_invert_matrix;
u8 *ecr_decode_matrix;
u8 *ecr_gf_tbls;
u8 *ecr_pp_gf_tbls;
};
static int
lov_ec_recover_init(struct ec_recover_ctx *ctx, struct lov_ec_rd_state *state,
int unavail_nr, bool reuse_state_bufs)
{
int dcount = state->ecr_dcount;
int pcount = state->ecr_pcount;
u8 *encode_matrix = state->ecr_encode_matrix;
int *err_array = state->ecr_err_array;
int *parity_used = state->ecr_parity_used;
u8 *temp_matrix = NULL;
u8 *invert_matrix = NULL;
u8 *decode_matrix = NULL;
u8 *tbls = NULL;
int i, j, k, r;
int rc = 0;
ENTRY;
/* temp_matrix/invert_matrix/decode_matrix are pure scratch space
* used only within this call to derive the GF tables, so both the
* RG-wide and per-page contexts can share one set, allocated into
* @state on-demand on first use. decode_matrix is sized for the
* worst case (pcount) since either context may need an unavail_nr
* up to pcount.
*/
if (!state->ecr_temp_matrix) {
OBD_ALLOC(state->ecr_temp_matrix, dcount * dcount);
if (!state->ecr_temp_matrix)
GOTO(out, rc = -ENOMEM);
}
if (!state->ecr_invert_matrix) {
OBD_ALLOC(state->ecr_invert_matrix, dcount * dcount);
if (!state->ecr_invert_matrix)
GOTO(out, rc = -ENOMEM);
}
if (!state->ecr_decode_matrix) {
OBD_ALLOC(state->ecr_decode_matrix, pcount * dcount);
if (!state->ecr_decode_matrix)
GOTO(out, rc = -ENOMEM);
}
temp_matrix = state->ecr_temp_matrix;
invert_matrix = state->ecr_invert_matrix;
decode_matrix = state->ecr_decode_matrix;
/* @ecr_gf_tbls (RG-wide) and @ecr_pp_gf_tbls (per-page) are kept
* independent: the RG-wide ctx must stay valid and unmodified
* across the whole recovery group for fast-path pages, while the
* per-page ctx is rebuilt for every slow-path page. Sharing one
* buffer would let a per-page rebuild silently corrupt the
* RG-wide tables.
*/
if (!state->ecr_gf_tbls) {
OBD_ALLOC(state->ecr_gf_tbls, dcount * pcount * 32);
if (!state->ecr_gf_tbls)
GOTO(out, rc = -ENOMEM);
}
if (!state->ecr_pp_gf_tbls) {
OBD_ALLOC(state->ecr_pp_gf_tbls, dcount * pcount * 32);
if (!state->ecr_pp_gf_tbls)
GOTO(out, rc = -ENOMEM);
}
if (!reuse_state_bufs)
tbls = state->ecr_gf_tbls;
else
tbls = state->ecr_pp_gf_tbls;
ctx->erc_dcount = dcount;
ctx->erc_pcount = pcount;
ctx->erc_unavail_nr = unavail_nr;
ctx->erc_tbls_size = dcount * pcount * 32;
ctx->erc_tbls = tbls;
/* get rid of coefficient rows of failed data stripes */
k = 0;
r = 0;
for (i = 0; i < dcount && r < dcount; i++, r++) {
while (k < unavail_nr && err_array[k] == r) {
k++;
r++;
}
for (j = 0; j < dcount; j++)
temp_matrix[i * dcount + j] =
encode_matrix[r * dcount + j];
}
/* append coefficient rows of used parity stripes */
if (parity_used) {
for (i = 0; i < unavail_nr; i++) {
int row = dcount + parity_used[i];
for (j = 0; j < dcount; j++)
temp_matrix[(dcount - unavail_nr + i) *
dcount + j] =
encode_matrix[row * dcount + j];
}
}
/* invert the matrix; a singular matrix here indicates a logic
* error in how err_array/parity_used were constructed
*/
rc = gf_invert_matrix(temp_matrix, invert_matrix, dcount);
if (rc < 0)
GOTO(out, rc = -EIO);
/* generate decode matrix by using only rows of failed data stripes
* from the inverted matrix
*/
for (i = 0; i < unavail_nr; i++) {
for (j = 0; j < dcount; j++)
decode_matrix[i * dcount + j] =
invert_matrix[err_array[i] * dcount + j];
}
ec_init_tables(dcount, unavail_nr, decode_matrix, ctx->erc_tbls);
rc = 0;
out:
/* The buffers are owned by state and will be freed in
* lov_ec_bufs_free(). We just need to handle error cleanup
* if allocation failed partway through.
*/
if (rc) {
ctx->erc_tbls = NULL;
if (!reuse_state_bufs) {
OBD_FREE(state->ecr_gf_tbls, dcount * pcount * 32);
state->ecr_gf_tbls = NULL;
}
OBD_FREE(state->ecr_pp_gf_tbls, dcount * pcount * 32);
state->ecr_pp_gf_tbls = NULL;
OBD_FREE(state->ecr_decode_matrix, pcount * dcount);
state->ecr_decode_matrix = NULL;
OBD_FREE(state->ecr_invert_matrix, dcount * dcount);
state->ecr_invert_matrix = NULL;
OBD_FREE(state->ecr_temp_matrix, dcount * dcount);
state->ecr_temp_matrix = NULL;
}
RETURN(rc);
}
static void
lov_ec_recover_page(struct ec_recover_ctx *ctx, unsigned char **ptrs,
unsigned char **recov_ptrs)
{
ec_encode_data(PAGE_SIZE, ctx->erc_dcount, ctx->erc_unavail_nr,
ctx->erc_tbls, ptrs, recov_ptrs);
}
/**
* lov_ec_bufs_free() - free per-cycle EC working buffers
* @state: state whose buffers to free
*
* Safe to call after lov_ec_bufs_alloc() whether it succeeded or not.
*/
static void lov_ec_bufs_free(struct lov_ec_rd_state *state)
{
unsigned int dcount = state->ecr_dcount;
unsigned int pcount = state->ecr_pcount;
OBD_FREE(state->ecr_gf_tbls, dcount * pcount * 32);
OBD_FREE(state->ecr_pp_gf_tbls, dcount * pcount * 32);
OBD_FREE(state->ecr_decode_matrix, pcount * dcount);
OBD_FREE(state->ecr_invert_matrix, dcount * dcount);
OBD_FREE(state->ecr_temp_matrix, dcount * dcount);
OBD_FREE(state->ecr_encode_matrix,
sizeof(*state->ecr_encode_matrix) *
dcount * (dcount + pcount));
OBD_FREE(state->ecr_zerobuf, sizeof(*state->ecr_zerobuf) * PAGE_SIZE);
OBD_FREE(state->ecr_recov_ptrs,
sizeof(*state->ecr_recov_ptrs) * pcount);
OBD_FREE(state->ecr_ptrs, sizeof(*state->ecr_ptrs) * dcount);
OBD_FREE(state->ecr_parity_used,
sizeof(*state->ecr_parity_used) * LOV_EC_MAX_CODING_STRIPES);
OBD_FREE(state->ecr_err_array,
sizeof(*state->ecr_err_array) * LOV_EC_MAX_CODING_STRIPES);
state->ecr_err_array = NULL;
state->ecr_parity_used = NULL;
state->ecr_ptrs = NULL;
state->ecr_recov_ptrs = NULL;
state->ecr_zerobuf = NULL;
state->ecr_encode_matrix = NULL;
state->ecr_temp_matrix = NULL;
state->ecr_invert_matrix = NULL;
state->ecr_decode_matrix = NULL;
state->ecr_gf_tbls = NULL;
state->ecr_pp_gf_tbls = NULL;
}
/**
* lov_ec_bufs_alloc() - allocate per-cycle EC working buffers
* @state: state to populate
* @dcount: data stripe count
* @pcount: parity stripe count (0 if non-EC component)
*
* Allocates err_array/parity_used (sized to LOV_EC_MAX_CODING_STRIPES),
* ptrs (dcount), and -- only when pcount > 0 -- recov_ptrs (pcount),
* zerobuf (PAGE_SIZE), and encode_matrix (dcount*(dcount+pcount)).
* Also generates the Cauchy encode matrix when pcount > 0.
*
* Return: 0 on success, -ENOMEM on failure (all partial allocations freed)
*/
static int lov_ec_bufs_alloc(struct lov_ec_rd_state *state,
unsigned int dcount, unsigned int pcount)
{
int rc = -ENOMEM;
ENTRY;
state->ecr_dcount = dcount;
state->ecr_pcount = pcount;
state->ecr_err_array = NULL;
state->ecr_parity_used = NULL;
state->ecr_ptrs = NULL;
state->ecr_recov_ptrs = NULL;
state->ecr_zerobuf = NULL;
state->ecr_encode_matrix = NULL;
state->ecr_temp_matrix = NULL;
state->ecr_invert_matrix = NULL;
state->ecr_decode_matrix = NULL;
state->ecr_gf_tbls = NULL;
state->ecr_pp_gf_tbls = NULL;
OBD_ALLOC(state->ecr_err_array,
sizeof(*state->ecr_err_array) * LOV_EC_MAX_CODING_STRIPES);
if (!state->ecr_err_array)
GOTO(out, rc);
OBD_ALLOC(state->ecr_parity_used,
sizeof(*state->ecr_parity_used) * LOV_EC_MAX_CODING_STRIPES);
if (!state->ecr_parity_used)
GOTO(out, rc);
OBD_ALLOC(state->ecr_ptrs, sizeof(*state->ecr_ptrs) * dcount);
if (!state->ecr_ptrs)
GOTO(out, rc);
if (pcount > 0) {
OBD_ALLOC(state->ecr_recov_ptrs,
sizeof(*state->ecr_recov_ptrs) * pcount);
if (!state->ecr_recov_ptrs)
GOTO(out, rc);
OBD_ALLOC(state->ecr_zerobuf,
sizeof(*state->ecr_zerobuf) * PAGE_SIZE);
if (!state->ecr_zerobuf)
GOTO(out, rc);
OBD_ALLOC(state->ecr_encode_matrix,
sizeof(*state->ecr_encode_matrix) * dcount *
(dcount + pcount));
if (!state->ecr_encode_matrix)
GOTO(out, rc);
gf_gen_cauchy1_matrix(state->ecr_encode_matrix,
dcount + pcount, dcount);
/* Recovery context buffers (temp_matrix, invert_matrix,
* decode_matrix, gf_tbls) are allocated on-demand in
* lov_ec_recover_init() when the RG-wide context is first
* built, then reused for per-page contexts. This avoids
* sleeping allocations inside the kmap loop while allowing
* buffer reuse.
*/
}
rc = 0;
out:
if (rc)
lov_ec_bufs_free(state);
RETURN(rc);
}
/**
* lov_ec_recover_page_pos() - reconstruct one page position in a recovery group
* @env: lu_env context
* @io: top-level cl_io
* @state: EC working buffers (ptrs, recov_ptrs, zerobuf, encode_matrix,
* err_array reused as per-page scratch, parity_used)
* @recover_ctx: shared recovery tables built for @err_nr full-stripe errors
* @all_pages: [npages * dcount] cl_page pointers, indexed [stripe*npages + j]
* @all_states: [npages * dcount] page states, same indexing
* @par_vmpages: [err_nr * npages] parity vmpages, indexed [pstripe*npages + j]
* @npages: number of page positions in this RG
* @j: page position to reconstruct
* @err_nr: number of error stripes for the shared ctx
* @dcount: data stripe count
* @pcount: parity stripe count
*
* For page position @j, classifies each stripe's page state, kmaps available
* data and parity pages, and calls lov_ec_recover_page() to reconstruct
* EC_DPG_ERROR pages from parity. Uses the shared @recover_ctx when the
* per-page error count (r_j) equals @err_nr (fast path), or builds a per-page
* ctx for the r_j actual errors (slow path, mixed cached+ERROR at position j).
*
* Return: 0 on success, negative errno on per-page recover_init failure.
* On failure, io->ci_ec.ec_recovery_failed is set so the caller can bail.
*/
static int lov_ec_recover_page_pos(const struct lu_env *env,
struct cl_io *io,
struct lov_ec_rd_state *state,
struct ec_recover_ctx *recover_ctx,
struct cl_page **all_pages,
enum ec_data_pg_state *all_states,
struct page **par_vmpages,
int npages, int j, int err_nr,
unsigned int dcount, unsigned int pcount)
{
struct ec_recover_ctx per_page_ctx = { 0 };
struct ec_recover_ctx *use_ctx = NULL;
int p_j = 0, r_j = 0;
int rc = 0;
int k;
for (k = 0; k < dcount; k++) {
struct page *vmpage = NULL;
int idx = k * npages + j;
if (all_pages[idx])
vmpage = all_pages[idx]->cp_vmpage;
switch (all_states[idx]) {
case EC_DPG_ZERO:
state->ecr_ptrs[p_j++] = state->ecr_zerobuf;
break;
case EC_DPG_UPTODATE:
case EC_DPG_READ:
if (vmpage)
state->ecr_ptrs[p_j++] =
kmap_local_page(vmpage);
break;
case EC_DPG_ERROR:
LASSERTF(r_j < pcount,
"error data index %d >= pcount %d\n",
r_j, pcount);
state->ecr_err_array[r_j] = k;
if (vmpage)
state->ecr_recov_ptrs[r_j] =
kmap_local_page(vmpage);
r_j++;
break;
}
}
/* Skip recovery if no errors, too many errors, or no parity pages
* available.
*/
if (r_j == 0 || r_j > pcount || !par_vmpages)
goto skip_recovery;
/* Fill parity ptrs from batch */
for (k = 0; k < r_j; k++) {
int pidx = k * npages + j;
state->ecr_ptrs[p_j + k] =
kmap_local_page(par_vmpages[pidx]);
}
if (r_j == err_nr) {
/* Fast path: err_array[0..r_j-1] == the original err_array[]
* exactly (same size, same sort order, same elements by
* construction).
*/
use_ctx = recover_ctx;
} else {
/* Slow path: mixed cached+ERROR at j. Build a per-page ctx
* for the r_j actual errors and the first r_j parity stripes.
* Reuse the RG-wide buffers already in state to avoid sleeping
* allocations while kmap_local mappings are live.
*/
rc = lov_ec_recover_init(&per_page_ctx, state, r_j, true);
if (rc == 0)
use_ctx = &per_page_ctx;
}
if (use_ctx)
lov_ec_recover_page(use_ctx, state->ecr_ptrs,
state->ecr_recov_ptrs);
/* kunmap parity in reverse order of mapping. kmap_local_page()
* mappings are stack-based and must be released in strict LIFO
* order (Documentation/mm/highmem.rst). The parity pages were
* mapped last, in increasing k, so unmap them in decreasing k.
*/
for (k = r_j - 1; k >= 0; k--) {
kunmap_local(state->ecr_ptrs[p_j + k]);
state->ecr_ptrs[p_j + k] = NULL;
}
skip_recovery:
/* kunmap data and recovery pages. These were acquired
* interleaved in the forward k loop above (available data pages
* into ecr_ptrs[], error data pages into ecr_recov_ptrs[]), so
* re-walk that loop in reverse and unmap each page as it is
* encountered, honouring the kmap_local LIFO contract. p_j and
* r_j are decremented to track the next slot exactly as the
* forward loop advanced them; the vmpage checks mirror the
* "if (vmpage)" guards that gated each kmap_local_page().
*/
for (k = (int)dcount - 1; k >= 0; k--) {
struct page *vmpage = NULL;
int idx = k * npages + j;
if (all_pages[idx])
vmpage = all_pages[idx]->cp_vmpage;
switch (all_states[idx]) {
case EC_DPG_ZERO:
p_j--;
break;
case EC_DPG_UPTODATE:
case EC_DPG_READ:
if (vmpage) {
p_j--;
kunmap_local(state->ecr_ptrs[p_j]);
state->ecr_ptrs[p_j] = NULL;
}
break;
case EC_DPG_ERROR:
r_j--;
if (vmpage) {
kunmap_local(state->ecr_recov_ptrs[r_j]);
state->ecr_recov_ptrs[r_j] = NULL;
}
break;
}
}
/* Per-page recover_init failed (-ENOMEM or -EIO from singular
* matrix): ERROR pages at positions 0..j are already in ec_page_list
* and will be disowned by vvp_io_ec_rd_end.
*/
if (rc)
io->u.ci_ec.ec_recovery_failed = true;
return rc;
}
static int lov_io_ec_rd_start(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *io = ios->cis_io;
struct lov_object *lov = lio->lis_object;
struct lov_device *dev = lov_object_dev(lov);
struct lov_layout_composite *comp = &lov->u.composite;
struct lov_layout_entry *lle_d = NULL; /* data component entry */
struct lov_layout_entry *lle_p = NULL; /* parity component entry */
struct lov_stripe_md_entry *lsme_d;
struct lov_stripe_md_entry *lsme_p;
int data_mir_idx; /* data mirror index */
int parity_mir_idx; /* parity mirror index */
int index; /* data component index */
int pindex; /* parity component index */
struct lov_ec_rd_state state = { 0 };
struct ec_recover_ctx recover_ctx = { 0 };
struct ec_split_comp sc = { 0 };
unsigned int dcount; /* erasure parameters: data stripe count */
unsigned int pcount; /* erasure parameters: parity stripe count */
unsigned int raid_set = 0; /* raid set number we are reading from */
unsigned long ss; /* stripe size */
unsigned int pgs_per_stripe; /* pages per stripe */
unsigned long RGs; /* recovery group size */
unsigned long long size; /* read size of this cycle */
int err_nr = 0; /* # of missing data chunks */
struct cl_page **all_pages = NULL;
enum ec_data_pg_state *all_states = NULL;
unsigned int array_count = 0;
struct cl_page **par_clpages = NULL;
struct page **par_vmpages = NULL;
unsigned int par_arr_cnt;
struct cl_page_list dlist;
struct cl_page_list plist;
struct cl_page *page;
int i, j, k;
int rc = 0;
ENTRY;
cl_page_list_init(&dlist);
cl_page_list_init(&plist);
/* VFS readahead can trigger EC recovery for pages beyond EOF. These
* pages have no data to recover - the VFS will zero-fill them - so
* just return success. Nothing past this raid set can have data
* either, so stop cl_io_loop() here instead of letting it walk every
* remaining raid set of the request (each cycle costs a getattr and
* a lock enqueue).
*/
if (lio->lis_pos >= io->u.ci_ec.ec_inode_size) {
io->u.ci_ec.ec_inner.crw_pos += io->u.ci_ec.ec_inner.crw_bytes;
io->u.ci_ec.ec_inner.crw_bytes = 0;
io->ci_continue = 0;
rc = lov_io_start(env, ios);
RETURN(rc);
}
data_mir_idx = lio->lis_mirror_index;
/* get data component */
index = lov_io_layout_at_mirror(lio, lio->lis_pos, data_mir_idx);
if (index < 0 || !lsm_entry_inited(lov->lo_lsm, index)) {
CWARN("%s: "DFID": failed to find valid data component covering file offset %llu, mirror: %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))), lio->lis_pos,
data_mir_idx, index);
RETURN(index < 0 ? index : -EINVAL);
}
parity_mir_idx = lov_parity_mirror_index_from_data(lio, data_mir_idx);
if (parity_mir_idx < 0) {
CWARN("%s: "DFID": failed to find valid parity mirror for data mirror %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))), data_mir_idx,
parity_mir_idx);
RETURN(parity_mir_idx);
}
lle_d = &comp->lo_entries[index];
/* [lis_pos, lis_endpos] could be a non-ec component, just data read
* with dcount <= stripe_count, else must be a ec component read
* lov_io_ec_rd_iter_init() has set each read cycle to include stripes
* from raid sets with the same EC parameters or with no EC.
*/
lsme_d = lle_d->lle_lsme;
dcount = lsme_d->lsme_dstripe_count;
if (dcount > 0) {
u64 offset;
u64 offset_i;
u64 row;
ec_split_stripes(lsme_d->lsme_stripe_count, dcount, &sc);
/* determine the raid set count and its stripe count */
offset = lio->lis_pos - lsme_d->lsme_extent.e_start;
row = div64_u64(offset, (u64)lsme_d->lsme_stripe_size *
lsme_d->lsme_stripe_count);
offset_i = div64_u64(offset, lsme_d->lsme_stripe_size) %
lsme_d->lsme_stripe_count;
if (offset_i < sc.esc_k0 * sc.esc_n0) {
dcount = sc.esc_k0;
raid_set = row * (sc.esc_n0 + sc.esc_n1) +
offset_i / sc.esc_k0;
} else {
dcount = sc.esc_k1;
raid_set = row * (sc.esc_n0 + sc.esc_n1) + sc.esc_n0 +
(offset_i - sc.esc_k0 * sc.esc_n0) /
sc.esc_k1;
}
pcount = lsme_d->lsme_cstripe_count;
} else {
/* with stripes with no EC, dcount == 0, set it to 4 */
if (lsme_d->lsme_stripe_count >= 4)
dcount = 4;
else
dcount = lsme_d->lsme_stripe_count;
pcount = 0;
}
size = min_t(__u64,
lsme_d->lsme_extent.e_end - lsme_d->lsme_extent.e_start,
lio->lis_endpos - lio->lis_pos);
ss = lsme_d->lsme_stripe_size;
LASSERTF(ss >= PAGE_SIZE,
"stripe size %lu is smaller than page size %lu\n",
ss, PAGE_SIZE);
/* size data to be read for each cycle, it's recovery group size for
* EC, or just data size for non EC read cycle
*/
RGs = ss * dcount;
pgs_per_stripe = (ss + PAGE_SIZE - 1) >> PAGE_SHIFT;
rc = lov_ec_bufs_alloc(&state, dcount, pcount);
if (rc)
GOTO(out, rc);
if (lsme_d->lsme_dstripe_count > 0) {
/* get parity component */
pindex = lov_io_layout_at_mirror(lio, lio->lis_pos,
parity_mir_idx);
if (pindex < 0 || !lsm_entry_inited(lov->lo_lsm, pindex)) {
CWARN("%s: "DFID": failed to find valid parity component covering file offset %llu, mirror: %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))), lio->lis_pos,
parity_mir_idx, pindex);
GOTO(out_ptrs, rc = pindex < 0 ? pindex : -EINVAL);
}
lle_p = &comp->lo_entries[pindex];
lsme_p = lle_p->lle_lsme;
}
CDEBUG(D_INODE, DFID": ec read comp %d (dcount %d pcount %d) from %llu to %llu, RGs %lu pgs_per_stripe %d\n",
PFID(lu_object_fid(lov2lu(lov))), index, dcount, pcount,
lio->lis_pos, lio->lis_pos + size, RGs, pgs_per_stripe);
/* read size is within the raid sets of the same EC parameters */
for (i = 0; i < DIV_ROUND_UP(size, RGs); i++, raid_set++) {
loff_t soff; /* start offset of a recovery group */
loff_t eoff; /* end offset of a recovery group */
unsigned int pg_seq_start;
unsigned int pg_seq_end;
unsigned int npages = 0;
int par_stripes_used = 0;
loff_t outer_start = io->u.ci_ec.ec_outer.crw_pos;
loff_t outer_end = io->u.ci_ec.ec_outer.crw_pos +
io->u.ci_ec.ec_outer.crw_bytes;
/* soff is at the beginning of a recovery group/raid set */
soff = lio->lis_pos + i * RGs;
eoff = soff + RGs;
if (eoff > lio->lis_pos + size)
eoff = lio->lis_pos + size;
/* each recovery groups contains dcount stripes, each page
* in these dcount stripes are calculated to get each page
* in the parity stripes.
*
* If outer read spans more than one stripe in a recovery group,
* we need to read the whole stripe.
*/
if (outer_start < soff)
outer_start = soff;
if (outer_end > eoff)
outer_end = eoff;
/* Read the whole stripe when the outer read spans more than
* one stripe within this RG (the per-stripe page ranges then
* differ between stripes, so only the full range is safe);
* otherwise read just the page positions the outer read needs.
*/
if (outer_end <= outer_start) {
pg_seq_start = 0;
pg_seq_end = 0;
} else if ((outer_start - soff) / ss !=
(outer_end - 1 - soff) / ss) {
pg_seq_start = 0;
pg_seq_end = pgs_per_stripe;
} else {
pg_seq_start = ((outer_start - soff) >> PAGE_SHIFT) %
pgs_per_stripe;
pg_seq_end = DIV_ROUND_UP(outer_end - soff, PAGE_SIZE) %
pgs_per_stripe;
if (pg_seq_end == 0)
pg_seq_end = pgs_per_stripe;
}
CDEBUG(D_INFO, DFID" read for page series from %d to %d\n",
PFID(lu_object_fid(lov2lu(lov))),
pg_seq_start, pg_seq_end);
/* Batched per-stripe reads for this RG */
npages = pg_seq_end - pg_seq_start;
array_count = npages * dcount;
OBD_ALLOC_PTR_ARRAY_LARGE(all_pages, array_count);
OBD_ALLOC_PTR_ARRAY_LARGE(all_states, array_count);
if (!all_pages || !all_states)
GOTO(out_err, rc = -ENOMEM);
/* Read all pages per stripe: one cl_io_submit_sync per
* stripe with all npages pages. Arrays indexed as
* [stripe * npages + page_pos].
*/
for (k = 0; k < dcount; k++) {
rc = lov_ec_read_stripe_pages(env, lio, index, k,
lsme_d, soff, eoff,
pg_seq_start, npages,
&all_pages[k * npages],
&all_states[k * npages],
&dlist);
if (rc) {
CWARN("%s: "DFID": fail to read stripe %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))),
k, rc);
io->u.ci_ec.ec_recovery_failed = true;
GOTO(rg_cleanup, rc);
}
}
/* Populate err_array from batch read states. A stripe with
* any ERROR page is unavailable.
*/
err_nr = 0;
for (k = 0; k < dcount; k++) {
int pg;
bool has_error = false;
for (pg = 0; pg < npages; pg++) {
if (all_states[k * npages + pg] ==
EC_DPG_ERROR) {
has_error = true;
break;
}
}
if (has_error)
state.ecr_err_array[err_nr++] = k;
if (err_nr >= LOV_EC_MAX_CODING_STRIPES ||
err_nr > pcount)
break;
}
/* Recheck after batch reads: per-page errors may exceed pcount
* even when the pre-check was OK (a second OST can die mid-IO,
* or all_states reveals more error stripes than lso_status
* predicted).
* Bail out cleanly so vvp_io_ec_rd_end skips the cache read
* and we return -EAGAIN instead of zeros.
*/
if (err_nr >= LOV_EC_MAX_CODING_STRIPES || err_nr > pcount) {
CWARN("%s: "DFID": too many unrecoverable stripes in raid set %d: err_nr %d > pcount %d: rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))),
raid_set, err_nr, pcount, -EIO);
io->u.ci_ec.ec_recovery_failed = true;
GOTO(rg_cleanup, rc = -EAGAIN);
}
/* Batch parity reads if needed. par_vmpages/par_clpages
* indexed as [parity_stripe * npages + page_pos].
*/
par_arr_cnt = npages * err_nr;
if (err_nr > 0 && err_nr <= pcount) {
unsigned long long poff;
int pi, ps;
poff = lsme_p->lsme_extent.e_start +
raid_set * (u64)pcount *
lsme_p->lsme_stripe_size;
OBD_ALLOC_PTR_ARRAY_LARGE(par_vmpages, par_arr_cnt);
OBD_ALLOC_PTR_ARRAY_LARGE(par_clpages, par_arr_cnt);
if (!par_vmpages || !par_clpages) {
io->u.ci_ec.ec_recovery_failed = true;
GOTO(rg_cleanup, rc = -ENOMEM);
}
/* Read parity stripes until we have enough (err_nr).
* Skip dead stripes.
*/
for (pi = 0; pi < pcount && par_stripes_used < err_nr;
pi++) {
ps = par_stripes_used;
rc = lov_ec_read_parity_stripe(env, lio, pindex,
pi, lsme_p, poff,
pg_seq_start, npages,
&par_vmpages[ps * npages],
&par_clpages[ps * npages],
&plist);
if (rc) {
rc = 0; /* skip, try next */
continue;
}
state.ecr_parity_used[par_stripes_used++] = pi;
}
if (par_stripes_used < err_nr) {
CWARN("%s: "DFID": not enough parity stripes (%d < %d): rc = %d\n",
lov2obd(dev->ld_lov)->obd_name,
PFID(lu_object_fid(lov2lu(lov))),
par_stripes_used, err_nr, -EIO);
/* plist cleanup handled by rg_cleanup: path */
io->u.ci_ec.ec_recovery_failed = true;
GOTO(rg_cleanup, rc = -EIO);
}
/* Init recovery tables; allocates working buffers
* into state for reuse by per-page contexts.
*/
rc = lov_ec_recover_init(&recover_ctx, &state, err_nr,
false);
if (rc) {
io->u.ci_ec.ec_recovery_failed = true;
/* plist cleanup handled by rg_cleanup: path */
GOTO(rg_cleanup, rc);
}
/* err_array has been used to build recover_ctx, it is
* fine to reuse it in the per-page error array below.
*/
}
/* Per-page reconstruction loop.
*
* The shared recover_ctx was built for err_nr unavailable
* stripes using err_array[]. At page position j, the number
* of actual EC_DPG_ERROR stripes (r_j) can be less than
* err_nr when some positions on an "error stripe" were
* already Uptodate in cache before recovery started. Using
* the shared ctx with r_j < err_nr would write past the end
* of recov_ptrs[] and crash in ec_encode_data_base.
*
* Build a per-page subset in err_array[0..r_j-1] (same sort
* order as the original err_array[]) and use the shared ctx
* only when r_j == err_nr (full-stripe failure, common case).
* Otherwise build a per-page ctx with the err_array[] subset
* and the first r_j entries of parity_used[] -- any r_j
* distinct parity stripes we already read are enough to
* recover r_j data positions. Reusing err_array[] is safe
* because recover_ctx already baked the original contents
* into erc_tbls.
*/
for (j = 0; j < npages; j++) {
rc = lov_ec_recover_page_pos(env, io, &state,
&recover_ctx, all_pages,
all_states, par_vmpages,
npages, j, err_nr,
dcount, pcount);
if (rc)
GOTO(rg_cleanup, rc);
} /* for each page position */
/* Mark all recovered pages Uptodate after the entire
* reconstruction loop completes. Doing this inside the
* per-page loop would mark pages for later positions
* Uptodate before they are actually reconstructed.
*/
if (!io->u.ci_ec.ec_recovery_failed) {
cl_page_list_for_each(page, &io->u.ci_ec.ec_page_list)
SetPageUptodate(cl_page_vmpage(page));
}
rg_cleanup:
/* release parity pages */
if (par_vmpages) {
for (k = 0; k < par_arr_cnt; k++) {
if (par_vmpages[k]) {
unlock_page(par_vmpages[k]);
__free_page(par_vmpages[k]);
}
}
}
/* the owernership of cl_page in par_clpages[] has been
* transferred to plist
*/
cl_page_list_fini(env, &plist);
OBD_FREE_PTR_ARRAY_LARGE(par_clpages, par_arr_cnt);
par_clpages = NULL;
OBD_FREE_PTR_ARRAY_LARGE(par_vmpages, par_arr_cnt);
par_vmpages = NULL;
/* Release data pages */
cl_page_list_disown(env, &dlist);
cl_page_list_fini(env, &dlist);
/* Release pages not in dlist(READ) nor in ec_page_list(ERROR):
* - UPTODATE pages: never added to any list
*/
for (k = 0; k < array_count; k++) {
if (all_pages[k]) {
if (all_states[k] == EC_DPG_UPTODATE) {
cl_page_disown(env, io, all_pages[k]);
cl_page_put(env, all_pages[k]);
}
all_pages[k] = NULL;
}
}
if (rc)
GOTO(out_err, rc);
OBD_FREE_PTR_ARRAY_LARGE(all_states, array_count);
all_states = NULL;
OBD_FREE_PTR_ARRAY_LARGE(all_pages, array_count);
all_pages = NULL;
/* batch scope */
} /* for each recovery group */
out_err:
OBD_FREE_PTR_ARRAY_LARGE(all_states, array_count);
OBD_FREE_PTR_ARRAY_LARGE(all_pages, array_count);
/* If bail out in the middle of recovery, we need to cleanup everything,
* including kunmap data and parity pages, release them.
*
* In practice these loops are defensive no-ops:
* lov_ec_recover_page_pos() is self-contained and kunmaps/NULLs every
* slot it maps before returning (both the parity loop and the
* skip_recovery data loop run unconditionally before its return),
* so by the time any path reaches out_err, ecr_ptrs[]/ecr_recov_ptrs[]
* are all NULL. The reverse iteration and per-slot NULL guards keep
* this safe should a future change bail with live mappings; a fully
* correct LIFO unwind in that case would need to re-walk
* all_pages[]/all_states[] exactly as lov_ec_recover_page_pos() does,
* since data and error-data mappings are interleaved in a single
* kmap_local stack.
*/
if (state.ecr_recov_ptrs)
for (k = (int)pcount - 1; k >= 0; k--)
if (state.ecr_recov_ptrs[k])
kunmap_local(state.ecr_recov_ptrs[k]);
if (state.ecr_ptrs)
for (k = (int)dcount - 1; k >= 0; k--)
if (state.ecr_ptrs[k] &&
state.ecr_ptrs[k] != state.ecr_zerobuf)
kunmap_local(state.ecr_ptrs[k]);
cl_page_list_disown(env, &dlist);
cl_page_list_fini(env, &dlist);
out_ptrs:
lov_ec_bufs_free(&state);
out:
if (rc != 0)
CDEBUG(D_INODE,
DFID": ec read comp %d failed: rc = %d\n",
PFID(lu_object_fid(lov2lu(lov))), index, rc);
/* inner io advance */
if (rc == 0) {
io->u.ci_ec.ec_inner.crw_pos += size;
io->u.ci_ec.ec_inner.crw_bytes -= size;
rc = lov_io_start(env, ios);
}
RETURN(rc);
}
static void
lov_io_data_version_end(const struct lu_env *env, const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *parent = lio->lis_cl.cis_io;
struct cl_data_version_io *pdv = &parent->u.ci_data_version;
struct lov_io_sub *sub;
ENTRY;
list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
struct cl_data_version_io *sdv = &sub->sub_io.u.ci_data_version;
lov_io_end_wrapper(sub->sub_env, &sub->sub_io);
pdv->dv_data_version += sdv->dv_data_version;
if (pdv->dv_layout_version > sdv->dv_layout_version)
pdv->dv_layout_version = sdv->dv_layout_version;
if (parent->ci_result == 0)
parent->ci_result = sub->sub_io.ci_result;
}
EXIT;
}
static void lov_io_iter_fini(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
int rc;
ENTRY;
OBD_FREE_PTR_ARRAY(lio->lis_trunc_stripe_index,
lio->lis_object->u.composite.lo_entry_count);
lio->lis_trunc_stripe_index = NULL;
rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
LASSERT(rc == 0);
while (!list_empty(&lio->lis_active))
list_del_init(lio->lis_active.next);
EXIT;
}
static void lov_io_unlock(const struct lu_env *env,
const struct cl_io_slice *ios)
{
int rc;
ENTRY;
rc = lov_io_call(env, cl2lov_io(env, ios), lov_io_unlock_wrapper);
LASSERT(rc == 0);
EXIT;
}
static int lov_io_read_ahead_prep(const struct lu_env *env,
const struct cl_io_slice *ios,
pgoff_t start, struct cl_read_ahead *ra)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct lov_object *loo = lio->lis_object;
struct lov_layout_raid0 *r0;
struct lov_io_sub *sub;
loff_t offset;
loff_t suboff;
pgoff_t ra_end;
unsigned int pps; /* pages per stripe */
int stripe;
int index;
int rc;
ENTRY;
offset = start << PAGE_SHIFT;
index = lov_io_layout_at(lio, offset);
if (index < 0 || !lsm_entry_inited(loo->lo_lsm, index) ||
lsm_entry_is_foreign(loo->lo_lsm, index))
RETURN(-ENODATA);
/* avoid readahead to expand to stale components */
if (!lov_entry(loo, index)->lle_valid)
RETURN(-EIO);
stripe = lov_stripe_number(loo->lo_lsm, index, offset);
r0 = lov_r0(loo, index);
if (unlikely(!r0->lo_sub[stripe]))
RETURN(-EIO);
sub = lov_sub_get(env, lio, lov_comp_index(index, stripe));
if (IS_ERR(sub))
RETURN(PTR_ERR(sub));
lov_stripe_offset(loo->lo_lsm, index, offset, stripe, &suboff);
rc = cl_io_read_ahead_prep(sub->sub_env, &sub->sub_io,
suboff >> PAGE_SHIFT, ra);
CDEBUG(D_READA, DFID " cra_end = %lu, stripes = %d, rc = %d\n",
PFID(lu_object_fid(lov2lu(loo))), ra->cra_end_idx,
r0->lo_nr, rc);
if (rc != 0)
RETURN(rc);
/**
* Adjust the stripe index by layout of comp. ra->cra_end is the
* maximum page index covered by an underlying DLM lock.
* This function converts cra_end from stripe level to file level, and
* make sure it's not beyond stripe and component boundary.
*/
/* cra_end is stripe level, convert it into file level */
ra_end = ra->cra_end_idx;
if (ra_end != CL_PAGE_EOF)
ra->cra_end_idx = lov_stripe_pgoff(loo->lo_lsm, index,
ra_end, stripe);
/* boundary of current component */
ra_end = lov_io_extent(lio, index)->e_end >> PAGE_SHIFT;
if (ra_end != CL_PAGE_EOF && ra->cra_end_idx >= ra_end)
ra->cra_end_idx = ra_end - 1;
if (r0->lo_nr == 1) /* single stripe file */
RETURN(0);
pps = lov_lse(loo, index)->lsme_stripe_size >> PAGE_SHIFT;
CDEBUG(D_READA, DFID " max_index = %lu, pps = %u, index = %d, stripe_size = %u, stripe no = %u, start index = %lu\n",
PFID(lu_object_fid(lov2lu(loo))), ra->cra_end_idx, pps, index,
lov_lse(loo, index)->lsme_stripe_size, stripe, start);
/* never exceed the end of the stripe */
ra->cra_end_idx = min_t(pgoff_t, ra->cra_end_idx,
start + pps - start % pps - 1);
RETURN(0);
}
static int lov_io_lru_reserve(const struct lu_env *env,
const struct cl_io_slice *ios, loff_t pos,
size_t bytes)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
struct lov_io_sub *sub;
struct lu_extent ext;
int index;
int rc = 0;
ENTRY;
ext.e_start = pos;
ext.e_end = pos + bytes;
lov_foreach_io_layout(index, lio, &ext) {
struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
struct lov_layout_raid0 *r0 = &le->lle_raid0;
u64 start;
u64 end;
int stripe;
if (!lsm_entry_inited(lsm, index))
continue;
if (!le->lle_valid && !ios->cis_io->ci_designated_mirror) {
CERROR(DFID": I/O to invalid component: %d, mirror: %d\n",
PFID(lu_object_fid(lov2lu(lio->lis_object))),
index, lio->lis_mirror_index);
RETURN(-EIO);
}
for (stripe = 0; stripe < r0->lo_nr; stripe++) {
if (!lov_stripe_intersects(lsm, index, stripe,
&ext, &start, &end))
continue;
if (unlikely(!r0->lo_sub[stripe]))
RETURN(-EIO);
sub = lov_sub_get(env, lio,
lov_comp_index(index, stripe));
if (IS_ERR(sub))
return PTR_ERR(sub);
rc = cl_io_lru_reserve(sub->sub_env, &sub->sub_io, start,
end - start + 1);
if (rc != 0)
RETURN(rc);
}
}
RETURN(0);
}
static int lov_dio_submit(const struct lu_env *env,
struct cl_io *io,
const struct cl_io_slice *ios,
enum cl_req_type crt, struct cl_dio_pages *cdp)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct lov_io_sub *sub;
int rc = 0;
int index;
ENTRY;
if (lov_pages_is_empty(cdp)) {
cl_dio_pages_complete(env, cdp, cdp->cdp_page_count, 0);
RETURN(0);
}
index = cdp->cdp_lov_index;
sub = lov_sub_get(env, lio, index);
if (!IS_ERR(sub)) {
rc = cl_dio_submit_rw(sub->sub_env, &sub->sub_io,
crt, cdp);
} else {
rc = PTR_ERR(sub);
}
RETURN(rc);
}
/**
* lov_io_submit() - lov implementation of cl_operations::cio_submit() method.
* @env: lustre execution environment
* @io: highlevel I/O request
* @ios: LOV specific IO
* @crt: Requested transfer type
* @queue: Page queue
*
* lov implementation of cl_operations::cio_submit() method. It takes a list
* of pages in @queue, splits it into per-stripe sub-lists, invokes
* cl_io_submit() on underlying devices to submit sub-lists, and then splices
* everything back.
*
* Major complication of this function is a need to handle memory cleansing:
* cl_io_submit() is called to write out pages as a part of VM memory
* reclamation, and hence it may not fail due to memory shortages (system
* dead-locks otherwise). To deal with this, some resources (sub-lists,
* sub-environment, etc.) are allocated per-device on "startup" (i.e., in a
* not-memory cleansing context), and in case of memory shortage, these
* pre-allocated resources are used by lov_io_submit() under
* lov_device::ld_mutex mutex.
*
* Return:
* * %0 on success
* * %negative on failure
*/
static int lov_io_submit(const struct lu_env *env,
struct cl_io *io,
const struct cl_io_slice *ios,
enum cl_req_type crt, struct cl_2queue *queue)
{
struct cl_page_list *qin = &queue->c2_qin;
struct lov_io *lio = cl2lov_io(env, ios);
struct lov_io_sub *sub;
struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
struct cl_page *page;
struct cl_page *tmp;
int index;
int rc = 0;
ENTRY;
cl_page_list_init(plist);
while (qin->pl_nr > 0) {
struct cl_2queue *cl2q = &lov_env_info(env)->lti_cl2q;
page = cl_page_list_first(qin);
if (lov_page_is_empty(page)) {
cl_page_list_move(&queue->c2_qout, qin, page);
/*
* it could only be mirror read to get here therefore
* the pages will be transient. We don't care about
* the return code of cl_page_prep() at all.
*/
LASSERT(page->cp_type == CPT_TRANSIENT);
cl_page_complete(env, page, crt, 0);
continue;
}
cl_2queue_init(cl2q);
cl_page_list_move(&cl2q->c2_qin, qin, page);
index = page->cp_lov_index;
cl_page_list_for_each_safe(page, tmp, qin) {
/* this page is not on this stripe */
if (index != page->cp_lov_index)
continue;
cl_page_list_move(&cl2q->c2_qin, qin, page);
}
sub = lov_sub_get(env, lio, index);
if (!IS_ERR(sub)) {
rc = cl_io_submit_rw(sub->sub_env, &sub->sub_io,
crt, cl2q);
} else {
rc = PTR_ERR(sub);
}
cl_page_list_splice(&cl2q->c2_qin, plist);
cl_page_list_splice(&cl2q->c2_qout, &queue->c2_qout);
cl_2queue_fini(env, cl2q);
if (rc != 0)
break;
}
cl_page_list_splice(plist, qin);
cl_page_list_fini(env, plist);
RETURN(rc);
}
static int lov_io_commit_async(const struct lu_env *env,
const struct cl_io_slice *ios,
struct cl_page_list *queue, int from, int to,
cl_commit_cbt cb, enum cl_io_priority prio)
{
struct cl_page_list *plist = &lov_env_info(env)->lti_plist;
struct lov_io *lio = cl2lov_io(env, ios);
bool hp = cl_io_high_prio(prio);
struct lov_io_sub *sub;
struct cl_page *page;
int rc = 0;
ENTRY;
if (lio->lis_nr_subios == 1) {
int idx = lio->lis_single_subio_index;
LASSERT(!lov_page_is_empty(cl_page_list_first(queue)));
sub = lov_sub_get(env, lio, idx);
LASSERT(!IS_ERR(sub));
LASSERT(sub == &lio->lis_single_subio);
rc = cl_io_commit_async(sub->sub_env, &sub->sub_io, queue,
from, to, cb, prio);
RETURN(rc);
}
cl_page_list_init(plist);
while (queue->pl_nr > 0) {
int stripe_to = to;
int index;
LASSERT(plist->pl_nr == 0);
page = cl_page_list_first(queue);
LASSERT(!lov_page_is_empty(page));
cl_page_list_move(plist, queue, page);
index = page->cp_lov_index;
while (queue->pl_nr > 0) {
page = cl_page_list_first(queue);
if (index != page->cp_lov_index)
break;
cl_page_list_move(plist, queue, page);
}
if (queue->pl_nr > 0) /* still has more pages */
stripe_to = PAGE_SIZE;
sub = lov_sub_get(env, lio, index);
if (!IS_ERR(sub)) {
rc = cl_io_commit_async(sub->sub_env, &sub->sub_io,
plist, from, stripe_to, cb,
prio);
} else {
rc = PTR_ERR(sub);
break;
}
if (plist->pl_nr > 0) /* short write */
break;
from = 0;
if (!hp && lov_comp_entry(index) !=
lov_comp_entry(page->cp_lov_index))
cl_io_extent_release(sub->sub_env, &sub->sub_io, prio);
}
if (rc == 0 && hp) {
list_for_each_entry(sub, &lio->lis_subios, sub_list)
cl_io_extent_release(sub->sub_env, &sub->sub_io, prio);
}
/* for error case, add the page back into the qin list */
LASSERT(ergo(rc == 0, plist->pl_nr == 0));
while (plist->pl_nr > 0) {
/* error occurred, add the uncommitted pages back into queue */
page = cl_page_list_last(plist);
cl_page_list_move_head(queue, plist, page);
}
RETURN(rc);
}
static int lov_io_fault_start(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct cl_fault_io *fio;
struct lov_io *lio;
struct lov_io_sub *sub;
loff_t offset;
int entry;
int stripe;
ENTRY;
fio = &ios->cis_io->u.ci_fault;
lio = cl2lov_io(env, ios);
/**
* LU-14502: ft_page could be an existing cl_page associated with
* the vmpage covering the fault index, and the page may still
* refer to another mirror of an old IO.
*/
if (lov_is_flr(lio->lis_object)) {
offset = fio->ft_index << PAGE_SHIFT;
entry = lov_io_layout_at(lio, offset);
if (entry < 0) {
CERROR(DFID": page fault index %lu invalid component: %d, mirror: %d\n",
PFID(lu_object_fid(&ios->cis_obj->co_lu)),
fio->ft_index, entry,
lio->lis_mirror_index);
RETURN(-EIO);
}
stripe = lov_stripe_number(lio->lis_object->lo_lsm,
entry, offset);
if (fio->ft_page->cp_lov_index !=
lov_comp_index(entry, stripe)) {
CDEBUG(D_INFO, DFID": page fault at index %lu, at mirror %u comp entry %u stripe %u, "
"been used with comp entry %u stripe %u\n",
PFID(lu_object_fid(&ios->cis_obj->co_lu)),
fio->ft_index, lio->lis_mirror_index,
entry, stripe,
lov_comp_entry(fio->ft_page->cp_lov_index),
lov_comp_stripe(fio->ft_page->cp_lov_index));
fio->ft_page->cp_lov_index =
lov_comp_index(entry, stripe);
}
}
sub = lov_sub_get(env, lio, fio->ft_page->cp_lov_index);
sub->sub_io.u.ci_fault.ft_bytes = fio->ft_bytes;
RETURN(lov_io_start(env, ios));
}
static int lov_io_setattr_start(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *parent = ios->cis_io;
struct lov_io_sub *sub;
struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
ENTRY;
if (cl_io_is_fallocate(parent)) {
list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
loff_t size = parent->u.ci_setattr.sa_attr.lvb_size;
int index = lov_comp_entry(sub->sub_subio_index);
int stripe = lov_comp_stripe(sub->sub_subio_index);
size = lov_size_to_stripe(lsm, index, size, stripe);
sub->sub_io.u.ci_setattr.sa_attr.lvb_size = size;
sub->sub_io.u.ci_setattr.sa_avalid =
parent->u.ci_setattr.sa_avalid;
}
}
RETURN(lov_io_start(env, ios));
}
static void lov_io_fsync_end(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct lov_io_sub *sub;
unsigned int *written = &ios->cis_io->u.ci_fsync.fi_nr_written;
ENTRY;
*written = 0;
list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
struct cl_io *subio = &sub->sub_io;
lov_io_end_wrapper(sub->sub_env, subio);
if (subio->ci_result == 0)
*written += subio->u.ci_fsync.fi_nr_written;
}
RETURN_EXIT;
}
static void lov_io_lseek_end(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *io = lio->lis_cl.cis_io;
struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
struct lov_io_sub *sub;
loff_t offset = -ENXIO;
__u64 hole_off = 0;
bool seek_hole = io->u.ci_lseek.ls_whence == SEEK_HOLE;
ENTRY;
list_for_each_entry(sub, &lio->lis_active, sub_linkage) {
struct cl_io *subio = &sub->sub_io;
int index = lov_comp_entry(sub->sub_subio_index);
int stripe = lov_comp_stripe(sub->sub_subio_index);
loff_t sub_off, lov_off;
__u64 comp_end = lsm->lsm_entries[index]->lsme_extent.e_end;
lov_io_end_wrapper(sub->sub_env, subio);
if (io->ci_result == 0)
io->ci_result = sub->sub_io.ci_result;
if (io->ci_result)
continue;
CDEBUG(D_INFO, DFID": entry %x stripe %u: SEEK_%s from %lld\n",
PFID(lu_object_fid(lov2lu(lio->lis_object))),
index, stripe, seek_hole ? "HOLE" : "DATA",
subio->u.ci_lseek.ls_start);
/* first subio with positive result is what we need */
sub_off = subio->u.ci_lseek.ls_result;
/* Expected error, offset is out of stripe file size */
if (sub_off == -ENXIO)
continue;
/* Any other errors are not expected with ci_result == 0 */
if (sub_off < 0) {
CDEBUG(D_INFO, "unexpected error: rc = %lld\n",
sub_off);
io->ci_result = sub_off;
continue;
}
lov_off = lov_stripe_size(lsm, index, sub_off + 1, stripe) - 1;
if (lov_off < 0) {
/* the only way to get negatove lov_off here is too big
* result. Return -EOVERFLOW then.
*/
io->ci_result = -EOVERFLOW;
CDEBUG(D_INFO, "offset %llu is too big: rc = %d\n",
(u64)lov_off, io->ci_result);
continue;
}
if (lov_off < io->u.ci_lseek.ls_start) {
io->ci_result = -EINVAL;
CDEBUG(D_INFO, "offset %lld < start %lld: rc = %d\n",
sub_off, io->u.ci_lseek.ls_start, io->ci_result);
continue;
}
/* resulting offset can be out of component range if stripe
* object is full and its file size was returned as virtual
* hole start. Skip this result, the next component will give
* us correct lseek result but keep possible hole offset in
* case there is no more components ahead
*/
if (lov_off >= comp_end) {
/* must be SEEK_HOLE case */
if (likely(seek_hole)) {
/* save comp end as potential hole offset */
hole_off = max_t(__u64, comp_end, hole_off);
} else {
io->ci_result = -EINVAL;
CDEBUG(D_INFO,
"off %lld >= comp_end %llu: rc = %d\n",
lov_off, comp_end, io->ci_result);
}
continue;
}
CDEBUG(D_INFO, "SEEK_%s: %lld->%lld/%lld: rc = %d\n",
seek_hole ? "HOLE" : "DATA",
subio->u.ci_lseek.ls_start, sub_off, lov_off,
sub->sub_io.ci_result);
offset = min_t(__u64, offset, lov_off);
}
/* no result but some component returns hole as component end */
if (seek_hole && offset == -ENXIO && hole_off > 0)
offset = hole_off;
io->u.ci_lseek.ls_result = offset;
RETURN_EXIT;
}
static const struct cl_io_operations lov_io_ops = {
.op = {
[CIT_READ] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_rw_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_start,
.cio_end = lov_io_end
},
[CIT_WRITE] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_rw_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_start,
.cio_end = lov_io_end
},
[CIT_EC_RD] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_ec_rd_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_ec_rd_start,
.cio_end = lov_io_end,
},
[CIT_SETATTR] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_setattr_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_setattr_start,
.cio_end = lov_io_end
},
[CIT_DATA_VERSION] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_start,
.cio_end = lov_io_data_version_end,
},
[CIT_FAULT] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_fault_start,
.cio_end = lov_io_end
},
[CIT_FSYNC] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_start,
.cio_end = lov_io_fsync_end
},
[CIT_LADVISE] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_start,
.cio_end = lov_io_end
},
[CIT_LSEEK] = {
.cio_fini = lov_io_fini,
.cio_iter_init = lov_io_iter_init,
.cio_iter_fini = lov_io_iter_fini,
.cio_lock = lov_io_lock,
.cio_unlock = lov_io_unlock,
.cio_start = lov_io_start,
.cio_end = lov_io_lseek_end
},
[CIT_GLIMPSE] = {
.cio_fini = lov_io_fini,
},
[CIT_MISC] = {
.cio_fini = lov_io_fini
}
},
.cio_read_ahead_prep = lov_io_read_ahead_prep,
.cio_lru_reserve = lov_io_lru_reserve,
.cio_submit = lov_io_submit,
.cio_dio_submit = lov_dio_submit,
.cio_commit_async = lov_io_commit_async,
};
/*
* Empty lov io operations.
*/
static void lov_empty_io_fini(const struct lu_env *env,
const struct cl_io_slice *ios)
{
struct lov_object *lov = cl2lov(ios->cis_obj);
struct lov_io *lio = cl2lov_io(env, ios);
struct cl_io *io = lio->lis_cl.cis_io;
ENTRY;
if (!(io->ci_type == CIT_MISC && io->ci_ignore_layout) &&
atomic_dec_and_test(&lov->lo_active_ios))
wake_up(&lov->lo_waitq);
EXIT;
}
static int lov_empty_dio_submit(const struct lu_env *env,
struct cl_io *io,
const struct cl_io_slice *ios,
enum cl_req_type crt, struct cl_dio_pages *cdp)
{
return -EBADF;
}
static int lov_empty_io_submit(const struct lu_env *env,
struct cl_io *io,
const struct cl_io_slice *ios,
enum cl_req_type crt, struct cl_2queue *queue)
{
return -EBADF;
}
static void lov_empty_impossible(const struct lu_env *env,
struct cl_io_slice *ios)
{
LBUG();
}
#define LOV_EMPTY_IMPOSSIBLE ((void *)lov_empty_impossible)
/*
* An io operation vector for files without stripes.
*/
static const struct cl_io_operations lov_empty_io_ops = {
.op = {
[CIT_READ] = {
.cio_fini = lov_empty_io_fini,
#if 0
.cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
.cio_lock = LOV_EMPTY_IMPOSSIBLE,
.cio_start = LOV_EMPTY_IMPOSSIBLE,
.cio_end = LOV_EMPTY_IMPOSSIBLE
#endif
},
[CIT_WRITE] = {
.cio_fini = lov_empty_io_fini,
.cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
.cio_lock = LOV_EMPTY_IMPOSSIBLE,
.cio_start = LOV_EMPTY_IMPOSSIBLE,
.cio_end = LOV_EMPTY_IMPOSSIBLE
},
[CIT_SETATTR] = {
.cio_fini = lov_empty_io_fini,
.cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
.cio_lock = LOV_EMPTY_IMPOSSIBLE,
.cio_start = LOV_EMPTY_IMPOSSIBLE,
.cio_end = LOV_EMPTY_IMPOSSIBLE
},
[CIT_FAULT] = {
.cio_fini = lov_empty_io_fini,
.cio_iter_init = LOV_EMPTY_IMPOSSIBLE,
.cio_lock = LOV_EMPTY_IMPOSSIBLE,
.cio_start = LOV_EMPTY_IMPOSSIBLE,
.cio_end = LOV_EMPTY_IMPOSSIBLE
},
[CIT_FSYNC] = {
.cio_fini = lov_empty_io_fini
},
[CIT_LADVISE] = {
.cio_fini = lov_empty_io_fini
},
[CIT_GLIMPSE] = {
.cio_fini = lov_empty_io_fini
},
[CIT_MISC] = {
.cio_fini = lov_empty_io_fini
}
},
.cio_submit = lov_empty_io_submit,
.cio_dio_submit = lov_empty_dio_submit,
.cio_commit_async = LOV_EMPTY_IMPOSSIBLE
};
int lov_io_init_composite(const struct lu_env *env, struct cl_object *obj,
struct cl_io *io)
{
struct lov_io *lio = lov_env_io(env);
struct lov_object *lov = cl2lov(obj);
int result;
ENTRY;
INIT_LIST_HEAD(&lio->lis_active);
result = lov_io_slice_init(env, lio, lov, io);
if (result)
GOTO(out, result);
result = lov_io_subio_init(env, lio, io);
if (!result) {
cl_io_slice_add(io, &lio->lis_cl, obj, &lov_io_ops);
if (!(io->ci_ignore_layout && io->ci_type == CIT_MISC))
atomic_inc(&lov->lo_active_ios);
}
EXIT;
out:
io->ci_result = result < 0 ? result : 0;
return result;
}
int lov_io_init_empty(const struct lu_env *env, struct cl_object *obj,
struct cl_io *io)
{
struct lov_object *lov = cl2lov(obj);
struct lov_io *lio = lov_env_io(env);
int result = 0;
ENTRY;
lio->lis_object = lov;
switch (io->ci_type) {
case CIT_MISC:
case CIT_GLIMPSE:
case CIT_READ:
case CIT_EC_RD:
result = 0;
break;
case CIT_FSYNC:
case CIT_LADVISE:
case CIT_LSEEK:
case CIT_SETATTR:
case CIT_DATA_VERSION:
result = +1;
break;
case CIT_WRITE:
result = -EBADF;
break;
case CIT_FAULT:
result = -EFAULT;
CERROR("Page fault on a file without stripes: "DFID"\n",
PFID(lu_object_fid(&obj->co_lu)));
break;
default:
LBUG();
}
if (result == 0) {
cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
if (!(io->ci_ignore_layout && io->ci_type == CIT_MISC))
atomic_inc(&lov->lo_active_ios);
}
io->ci_result = result < 0 ? result : 0;
RETURN(result);
}
int lov_io_init_released(const struct lu_env *env, struct cl_object *obj,
struct cl_io *io)
{
struct lov_object *lov = cl2lov(obj);
struct lov_io *lio = lov_env_io(env);
int result;
ENTRY;
LASSERT(lov->lo_lsm != NULL);
lio->lis_object = lov;
switch (io->ci_type) {
default:
LASSERTF(0, "invalid type %d\n", io->ci_type);
result = -EOPNOTSUPP;
break;
case CIT_GLIMPSE:
case CIT_MISC:
case CIT_FSYNC:
case CIT_LADVISE:
case CIT_DATA_VERSION:
result = 1;
break;
case CIT_SETATTR:
/*
* the truncate to 0 is managed by MDT:
* - in open, for open O_TRUNC
* - in setattr, for truncate
*/
/*
* the truncate is for size > 0 so triggers a restore,
* also trigger a restore for prealloc/punch
*/
if (cl_io_is_trunc(io) || cl_io_is_fallocate(io)) {
io->ci_restore_needed = 1;
result = -ENODATA;
} else
result = 1;
break;
case CIT_READ:
case CIT_WRITE:
case CIT_FAULT:
case CIT_LSEEK:
case CIT_EC_RD:
io->ci_restore_needed = 1;
result = -ENODATA;
break;
}
if (result == 0) {
cl_io_slice_add(io, &lio->lis_cl, obj, &lov_empty_io_ops);
if (!(io->ci_ignore_layout && io->ci_type == CIT_MISC))
atomic_inc(&lov->lo_active_ios);
}
io->ci_result = result < 0 ? result : 0;
RETURN(result);
}
/* confirm this offset is in the given layout entry */
bool lov_io_layout_at_confirm(struct lov_io *lio, int entry, __u64 offset)
{
struct lov_object *lov = lio->lis_object;
struct lov_layout_entry *lle = lov_entry(lov, entry);
if ((offset >= lle->lle_extent->e_start &&
offset < lle->lle_extent->e_end) ||
(offset == OBD_OBJECT_EOF &&
lle->lle_extent->e_end == OBD_OBJECT_EOF))
return true;
return false;
}
/**
* lov_io_layout_at_mirror() - Return the index in composite layout given
* file offset
*
* @lio: Pointer to struct lov_io
* @offset: Offset in the composite layout
* @mirror_idx: for FLR, index of the mirror where to find the component
*
* Return the index in composite:lo_entries by the file offset
*
* Return:
* * %>=0 composite index where offset is located
* * %negative on error
*/
int lov_io_layout_at_mirror(struct lov_io *lio, __u64 offset, int mirror_index)
{
struct lov_object *lov = lio->lis_object;
struct lov_layout_composite *comp = &lov->u.composite;
int start_index = 0;
int end_index = comp->lo_entry_count - 1;
int i;
LASSERT(lov->lo_type == LLT_COMP);
/* This is actual file offset so nothing can cover eof. */
if (offset == LUSTRE_EOF)
return -EINVAL;
if (lov_is_flr(lov)) {
struct lov_mirror_entry *lre;
if (mirror_index < 0 ||
mirror_index >= lov->u.composite.lo_mirror_count)
return -EINVAL;
lre = lov_mirror_entry(lov, mirror_index);
start_index = lre->lre_start;
end_index = lre->lre_end;
}
for (i = start_index; i <= end_index; i++) {
struct lov_layout_entry *lle = lov_entry(lov, i);
LASSERT(!lsme_is_foreign(lle->lle_lsme));
if ((offset >= lle->lle_extent->e_start &&
offset < lle->lle_extent->e_end) ||
(offset == OBD_OBJECT_EOF &&
lle->lle_extent->e_end == OBD_OBJECT_EOF))
return i;
}
return -EINVAL;
}
int lov_io_layout_at(struct lov_io *lio, __u64 offset)
{
return lov_io_layout_at_mirror(lio, offset, lio->lis_mirror_index);
}
/** @} lov */
