Place a memory barrier around cp_state changes
PATCH 1/1 Place a memory barrier around cp_state changes
Ensure all uses of cl_page->cp_state are smp-safe.
LBUG Output:
In __cl_page_state_set()
..
old = page->cp_state;
PASSERT(env, page, allowed_transitions[old][state]);
..
Asserted with the following:
page at ffff80be1fcd6600[3 ffff80be1fca5cc0 0 1 ffff809e60e2 8cc0]
However cp_state 0 (CPS_CACHED) to 1 (CPS_OWNED) is a valid transition leading to the conclusion that cp_state became 0 during the assertion.
--- fs/lustre/include/cl_object.h | 11 +++++++++++ fs/lustre/llite/rw26.c | 2 +- fs/lustre/llite/vvp_page.c | 6 +++--- fs/lustre/obdclass/cl_page.c | 34 +++++++++++++++++++-------------- fs/lustre/obdecho/echo_client.c | 2 +- fs/lustre/osc/osc_cache.c | 18 +++++++++-------- 6 files changed, 46 insertions(+), 27 deletions(-)
fs/lustre/include/cl_object.h +11 -0
@@ -752,6 +752,17 @@ struct cl_page {struct cl_sync_io *cp_sync_io; }; +static inline enum cl_page_state cl_page_state_get (const struct cl_page *pg) +{ +/* +* Paired with smp_store_release in cl_page_state_set_trust +* and ensures that we see the most recent value of cp_state +* even when the last modification was not performed on the +* current processor +*/ +return smp_load_acquire (&pg->cp_state); +} +/** * Per-layer part of cl_page. *
fs/lustre/llite/rw26.c +1 -1
@@ -200,7 +200,7 @@ static ssize_t ll_direct_IO_seg(const struct lu_env *env, struct cl_io *io, rc =cl_page_own (env, io, clp);if (rc) { -LASSERT (clp->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (clp) ==CPS_FREEING );cl_page_put (env, clp);break ; }
fs/lustre/llite/vvp_page.c +3 -3
@@ -323,18 +323,18 @@ static int vvp_page_make_ready(const struct lu_env *env,lock_page (vmpage);if (clear_page_dirty_for_io (vmpage)) { -LASSERT (pg->cp_state ==CPS_CACHED ); +LASSERT (cl_page_state_get (pg) ==CPS_CACHED );/* This actually clears the dirty bit in the radix tree. */ set_page_writeback (vmpage);CL_PAGE_HEADER (D_PAGE , env, pg,"readied\n" ); - }else if (pg->cp_state ==CPS_PAGEOUT ) { + }else if (cl_page_state_get (pg) ==CPS_PAGEOUT ) {/* is it possible for osc_flush_async_page() to already * make it ready? */ result = -EALREADY ; }else {CL_PAGE_DEBUG (D_ERROR , env, pg,"Unexpecting page state %d.\n" , - pg->cp_state); +cl_page_state_get (pg));LBUG (); }unlock_page (vmpage);
fs/lustre/obdclass/cl_page.c +20 -14
@@ -97,7 +97,7 @@ static void cl_page_free(const struct lu_env *env, struct cl_page *page)PASSERT (env, page,list_empty (&page->cp_batch));PASSERT (env, page, !page->cp_owner); -PASSERT (env, page, page->cp_state ==CPS_FREEING ); +PASSERT (env, page,cl_page_state_get (page) ==CPS_FREEING ); while ((slice = list_first_entry_or_null(&page->cp_layers, struct cl_page_slice,
@@ -119,8 +119,14 @@ static void cl_page_free(const struct lu_env *env, struct cl_page *page)static inline void cl_page_state_set_trust (struct cl_page *page,enum cl_page_state state) { -/* bypass const. */ - *(enum cl_page_state *)&page->cp_state = state; +/* +* Paired with smp_load_acquire in cl_page_state_get +* and ensures that we see the most recent value of cp_state +* is available even when the next access is not performed on the +* current processor. +* Note we also cast away const as the only modifier of cp_state. +*/ +smp_store_release ((enum cl_page_state *)&page->cp_state, state); } struct cl_page *cl_page_alloc(const struct lu_env *env,
@@ -270,10 +276,10 @@ static void __cl_page_state_set(const struct lu_env *env, } }; - old = page->cp_state; + old =cl_page_state_get (page);PASSERT (env, page, allowed_transitions[old][state]);CL_PAGE_HEADER (D_TRACE , env, page,"%d -> %d\n" , old, state); -PASSERT (env, page, page->cp_state == old); +PASSERT (env, page,cl_page_state_get (page) == old);PASSERT (env, page,equi (state ==CPS_OWNED , page->cp_owner));cl_page_state_set_trust (page, state); }
@@ -313,7 +319,7 @@ void cl_page_put(const struct lu_env *env, struct cl_page *page) refcount_read(&page->cp_ref));if (refcount_dec_and_test (&page->cp_ref)) { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING );LASSERT (refcount_read (&page->cp_ref) ==0 );PASSERT (env, page, !page->cp_owner);
@@ -378,7 +384,7 @@ void __cl_page_disown(const struct lu_env *env,const struct cl_page_slice *slice;enum cl_page_state state; - state = pg->cp_state; + state =cl_page_state_get (pg);cl_page_owner_clear (pg);if (state ==CPS_OWNED )
@@ -402,7 +408,7 @@ int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)struct cl_io *top =cl_io_top ((struct cl_io *)io);LINVRNT (cl_object_same (pg->cp_obj, io->ci_obj)); -return pg->cp_state ==CPS_OWNED && pg->cp_owner == top; +return cl_page_state_get (pg) ==CPS_OWNED && pg->cp_owner == top; }EXPORT_SYMBOL (cl_page_is_owned);
@@ -434,7 +440,7 @@ static int __cl_page_own(const struct lu_env *env, struct cl_io *io, io =cl_io_top (io); -if (pg->cp_state ==CPS_FREEING ) { +if (cl_page_state_get (pg) ==CPS_FREEING ) { result = -ENOENT ; goto out; }
@@ -453,7 +459,7 @@ static int __cl_page_own(const struct lu_env *env, struct cl_io *io,PASSERT (env, pg, !pg->cp_owner); pg->cp_owner =cl_io_top (io);cl_page_owner_set (pg); -if (pg->cp_state !=CPS_FREEING ) { +if (cl_page_state_get (pg) !=CPS_FREEING ) {cl_page_state_set (env, pg,CPS_OWNED ); }else {__cl_page_disown (env, io, pg);
@@ -593,7 +599,7 @@ static void __cl_page_delete(const struct lu_env *env, struct cl_page *pg) {const struct cl_page_slice *slice; -PASSERT (env, pg, pg->cp_state !=CPS_FREEING ); +PASSERT (env, pg,cl_page_state_get (pg) !=CPS_FREEING );/* * Sever all ways to obtain new pointers to @pg.
@@ -756,7 +762,7 @@ void cl_page_completion(const struct lu_env *env,const struct cl_page_slice *slice;PASSERT (env, pg, crt <CRT_NR ); -PASSERT (env, pg, pg->cp_state ==cl_req_type_state (crt)); +PASSERT (env, pg,cl_page_state_get (pg) ==cl_req_type_state (crt));CL_PAGE_HEADER (D_TRACE , env, pg,"%d %d\n" , crt, ioret);
@@ -805,7 +811,7 @@ int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg, }if (result >=0 ) { -PASSERT (env, pg, pg->cp_state ==CPS_CACHED ); +PASSERT (env, pg,cl_page_state_get (pg) ==CPS_CACHED );cl_page_io_start (env, pg, crt); result =0 ; }
@@ -870,7 +876,7 @@ void cl_page_header_print(const struct lu_env *env, void *cookie, (*printer)(env, cookie,"page@%p[%d %p %d %d %p]\n" , pg,refcount_read (&pg->cp_ref), pg->cp_obj, - pg->cp_state, pg->cp_type, +cl_page_state_get (pg), pg->cp_type, pg->cp_owner); }EXPORT_SYMBOL (cl_page_header_print);
fs/lustre/obdecho/echo_client.c +1 -1
@@ -1046,7 +1046,7 @@ static int cl_echo_object_brw(struct echo_object *eco, int rw, u64 offset, rc =cl_page_own (env, io, clp);if (rc) { -LASSERT (clp->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (clp) ==CPS_FREEING );cl_page_put (env, clp);break ; }
fs/lustre/osc/osc_cache.c +10 -8
@@ -1045,7 +1045,7 @@ static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING );LASSERT (0 ); }
@@ -1356,10 +1356,12 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,int srvlock; cmd &= ~OBD_BRW_NOQUOTA ; -LASSERTF (equi (page->cp_state ==CPS_PAGEIN , cmd ==OBD_BRW_READ ), -"cp_state:%u, cmd:%d\n" , page->cp_state, cmd); -LASSERTF (equi (page->cp_state ==CPS_PAGEOUT , cmd ==OBD_BRW_WRITE ), -"cp_state:%u, cmd:%d\n" , page->cp_state, cmd); +LASSERTF (equi (cl_page_state_get (page) ==CPS_PAGEIN , + cmd ==OBD_BRW_READ ), +"cp_state:%u, cmd:%d\n" ,cl_page_state_get (page), cmd); +LASSERTF (equi (cl_page_state_get (page) ==CPS_PAGEOUT , + cmd ==OBD_BRW_WRITE ), +"cp_state:%u, cmd:%d\n" ,cl_page_state_get (page), cmd);LASSERT (opg->ops_transfer_pinned); crt = cmd ==OBD_BRW_READ ?CRT_READ :CRT_WRITE ;
@@ -3061,7 +3063,7 @@ bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io, page = ops->ops_cl.cpl_page;LASSERT (page->cp_type ==CPT_CACHEABLE ); -if (page->cp_state ==CPS_FREEING ) +if (cl_page_state_get (page) ==CPS_FREEING )continue ;cl_page_get (page);
@@ -3142,7 +3144,7 @@ static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING ); } }
@@ -3169,7 +3171,7 @@ static bool discard_cb(const struct lu_env *env, struct cl_io *io,cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING ); }return true;
Signature · 1 line
2.17.1
PATCH 1/1 Place a memory barrier around cp_state changes
Quoted 10 lines · Shaun Tancheff
Ensure all uses of
cl_page->cp_stateare smp-safe.LBUG Output:
In__cl_page_state_set()
..
old =page->cp_state;
PASSERT(env, page, allowed_transitions[old][state]);
..
Asserted with the following:
page at ffff80be1fcd6600[3ffff80be1fca5cc00 1ffff809e60e28cc0]
What do all these fields mean? Or where is this printed? The closest I can find is in cl_page_header_print() which uses the format string:
"page@%p[%d %p %d %d %p]\n",
But that only has 5 fields in the [], while the output you provided has 6.
Quoted 4 lines
However
cp_state0 (CPS_CACHED) to 1 (CPS_OWNED) is a valid transition leading to the conclusion thatcp_statebecame 0 during the assertion.
Where is the evidence that this was the transition that was happening? If it was in some part of the LBUG output that wasn't quoted - then please fix that by quoting the entire LBUG output.
We need to understand which change happened at this time to cause the race.
Then we need to explain why the added barriers actually close the race window.
If these barrier do actually fix a race, then there is something *very* wrong.
The comments say that changes to cp_state are protected by the page lock on the corresponding VM page.
Locking and unlocking a VM page entails sufficient barriers that changes made while one thread holds the lock will be visible to another thread once it also gets the lock.
So the only possible explanation for a race as you suspect, is if the vmpage *isn't* locked when something changes cp_state, and that would be bad.
Thanks,
NeilBrown
Quoted 268 lines
Signed-off-by Shaun Tancheff <[email protected]>--- fs/lustre/include/cl_object.h | 11 +++++++++++ fs/lustre/llite/rw26.c | 2 +- fs/lustre/llite/vvp_page.c | 6 +++--- fs/lustre/obdclass/cl_page.c | 34 +++++++++++++++++++-------------- fs/lustre/obdecho/echo_client.c | 2 +- fs/lustre/osc/osc_cache.c | 18 +++++++++-------- 6 files changed, 46 insertions(+), 27 deletions(-)fs/lustre/include/cl_object.h +11
-0@@ -752,6 +752,17 @@ struct cl_page {struct cl_sync_io *cp_sync_io; }; +static inline enum cl_page_state cl_page_state_get (const struct cl_page *pg) +{ +/* +* Paired with smp_store_release in cl_page_state_set_trust +* and ensures that we see the most recent value of cp_state +* even when the last modification was not performed on the +* current processor +*/ +return smp_load_acquire (&pg->cp_state); +} +/** * Per-layer part of cl_page. * fs/lustre/llite/rw26.c +1
-1@@ -200,7 +200,7 @@ static ssize_t ll_direct_IO_seg(const struct lu_env *env, struct cl_io *io, rc =cl_page_own (env, io, clp);if (rc) { -LASSERT (clp->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (clp) ==CPS_FREEING );cl_page_put (env, clp);break ; }fs/lustre/llite/vvp_page.c +3
-3@@ -323,18 +323,18 @@ static int vvp_page_make_ready(const struct lu_env *env,lock_page (vmpage);if (clear_page_dirty_for_io (vmpage)) { -LASSERT (pg->cp_state ==CPS_CACHED ); +LASSERT (cl_page_state_get (pg) ==CPS_CACHED );/* This actually clears the dirty bit in the radix tree. */ set_page_writeback (vmpage);CL_PAGE_HEADER (D_PAGE , env, pg,"readied\n" ); - }else if (pg->cp_state ==CPS_PAGEOUT ) { + }else if (cl_page_state_get (pg) ==CPS_PAGEOUT ) {/* is it possible for osc_flush_async_page() to already * make it ready? */ result = -EALREADY ; }else {CL_PAGE_DEBUG (D_ERROR , env, pg,"Unexpecting page state %d.\n" , - pg->cp_state); +cl_page_state_get (pg));LBUG (); }unlock_page (vmpage);fs/lustre/obdclass/cl_page.c +20
-14@@ -97,7 +97,7 @@ static void cl_page_free(const struct lu_env *env, struct cl_page *page)PASSERT (env, page,list_empty (&page->cp_batch));PASSERT (env, page, !page->cp_owner); -PASSERT (env, page, page->cp_state ==CPS_FREEING ); +PASSERT (env, page,cl_page_state_get (page) ==CPS_FREEING ); while ((slice = list_first_entry_or_null(&page->cp_layers, struct cl_page_slice,@@ -119,8 +119,14 @@ static void cl_page_free(const struct lu_env *env, struct cl_page *page)static inline void cl_page_state_set_trust (struct cl_page *page,enum cl_page_state state) { -/* bypass const. */ - *(enum cl_page_state *)&page->cp_state = state; +/* +* Paired with smp_load_acquire in cl_page_state_get +* and ensures that we see the most recent value of cp_state +* is available even when the next access is not performed on the +* current processor. +* Note we also cast away const as the only modifier of cp_state. +*/ +smp_store_release ((enum cl_page_state *)&page->cp_state, state); } struct cl_page *cl_page_alloc(const struct lu_env *env,@@ -270,10 +276,10 @@ static void __cl_page_state_set(const struct lu_env *env, } }; - old = page->cp_state; + old =cl_page_state_get (page);PASSERT (env, page, allowed_transitions[old][state]);CL_PAGE_HEADER (D_TRACE , env, page,"%d -> %d\n" , old, state); -PASSERT (env, page, page->cp_state == old); +PASSERT (env, page,cl_page_state_get (page) == old);PASSERT (env, page,equi (state ==CPS_OWNED , page->cp_owner));cl_page_state_set_trust (page, state); }@@ -313,7 +319,7 @@ void cl_page_put(const struct lu_env *env, struct cl_page *page) refcount_read(&page->cp_ref));if (refcount_dec_and_test (&page->cp_ref)) { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING );LASSERT (refcount_read (&page->cp_ref) ==0 );PASSERT (env, page, !page->cp_owner);@@ -378,7 +384,7 @@ void __cl_page_disown(const struct lu_env *env,const struct cl_page_slice *slice;enum cl_page_state state; - state = pg->cp_state; + state =cl_page_state_get (pg);cl_page_owner_clear (pg);if (state ==CPS_OWNED )@@ -402,7 +408,7 @@ int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)struct cl_io *top =cl_io_top ((struct cl_io *)io);LINVRNT (cl_object_same (pg->cp_obj, io->ci_obj)); -return pg->cp_state ==CPS_OWNED && pg->cp_owner == top; +return cl_page_state_get (pg) ==CPS_OWNED && pg->cp_owner == top; }EXPORT_SYMBOL (cl_page_is_owned);@@ -434,7 +440,7 @@ static int __cl_page_own(const struct lu_env *env, struct cl_io *io, io =cl_io_top (io); -if (pg->cp_state ==CPS_FREEING ) { +if (cl_page_state_get (pg) ==CPS_FREEING ) { result = -ENOENT ; goto out; }@@ -453,7 +459,7 @@ static int __cl_page_own(const struct lu_env *env, struct cl_io *io,PASSERT (env, pg, !pg->cp_owner); pg->cp_owner =cl_io_top (io);cl_page_owner_set (pg); -if (pg->cp_state !=CPS_FREEING ) { +if (cl_page_state_get (pg) !=CPS_FREEING ) {cl_page_state_set (env, pg,CPS_OWNED ); }else {__cl_page_disown (env, io, pg);@@ -593,7 +599,7 @@ static void __cl_page_delete(const struct lu_env *env, struct cl_page *pg) {const struct cl_page_slice *slice; -PASSERT (env, pg, pg->cp_state !=CPS_FREEING ); +PASSERT (env, pg,cl_page_state_get (pg) !=CPS_FREEING );/* * Sever all ways to obtain new pointers to @pg. @@ -756,7 +762,7 @@ void cl_page_completion(const struct lu_env *env,const struct cl_page_slice *slice;PASSERT (env, pg, crt <CRT_NR ); -PASSERT (env, pg, pg->cp_state ==cl_req_type_state (crt)); +PASSERT (env, pg,cl_page_state_get (pg) ==cl_req_type_state (crt));CL_PAGE_HEADER (D_TRACE , env, pg,"%d %d\n" , crt, ioret);@@ -805,7 +811,7 @@ int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg, }if (result >=0 ) { -PASSERT (env, pg, pg->cp_state ==CPS_CACHED ); +PASSERT (env, pg,cl_page_state_get (pg) ==CPS_CACHED );cl_page_io_start (env, pg, crt); result =0 ; }@@ -870,7 +876,7 @@ void cl_page_header_print(const struct lu_env *env, void *cookie, (*printer)(env, cookie,"page@%p[%d %p %d %d %p]\n" , pg,refcount_read (&pg->cp_ref), pg->cp_obj, - pg->cp_state, pg->cp_type, +cl_page_state_get (pg), pg->cp_type, pg->cp_owner); }EXPORT_SYMBOL (cl_page_header_print);fs/lustre/obdecho/echo_client.c +1
-1@@ -1046,7 +1046,7 @@ static int cl_echo_object_brw(struct echo_object *eco, int rw, u64 offset, rc =cl_page_own (env, io, clp);if (rc) { -LASSERT (clp->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (clp) ==CPS_FREEING );cl_page_put (env, clp);break ; }fs/lustre/osc/osc_cache.c +10
-8@@ -1045,7 +1045,7 @@ static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING );LASSERT (0 ); }@@ -1356,10 +1356,12 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,int srvlock; cmd &= ~OBD_BRW_NOQUOTA ; -LASSERTF (equi (page->cp_state ==CPS_PAGEIN , cmd ==OBD_BRW_READ ), -"cp_state:%u, cmd:%d\n" , page->cp_state, cmd); -LASSERTF (equi (page->cp_state ==CPS_PAGEOUT , cmd ==OBD_BRW_WRITE ), -"cp_state:%u, cmd:%d\n" , page->cp_state, cmd); +LASSERTF (equi (cl_page_state_get (page) ==CPS_PAGEIN , + cmd ==OBD_BRW_READ ), +"cp_state:%u, cmd:%d\n" ,cl_page_state_get (page), cmd); +LASSERTF (equi (cl_page_state_get (page) ==CPS_PAGEOUT , + cmd ==OBD_BRW_WRITE ), +"cp_state:%u, cmd:%d\n" ,cl_page_state_get (page), cmd);LASSERT (opg->ops_transfer_pinned); crt = cmd ==OBD_BRW_READ ?CRT_READ :CRT_WRITE ;@@ -3061,7 +3063,7 @@ bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io, page = ops->ops_cl.cpl_page;LASSERT (page->cp_type ==CPT_CACHEABLE ); -if (page->cp_state ==CPS_FREEING ) +if (cl_page_state_get (page) ==CPS_FREEING )continue ;cl_page_get (page);@@ -3142,7 +3144,7 @@ static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING ); } }@@ -3169,7 +3171,7 @@ static bool discard_cb(const struct lu_env *env, struct cl_io *io,cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING ); }return true;Signature · 1 line
2.17.1
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20190607/c70353a9/attachment.sig>
PATCH 1/1 Place a memory barrier around cp_state changes
Quoted 14 lines · NeilBrown
Quoted 10 lines · Shaun Tancheff
Ensure all uses of
cl_page->cp_stateare smp-safe.LBUG Output:
In__cl_page_state_set()
..
old =page->cp_state;
PASSERT(env, page, allowed_transitions[old][state]);
..
Asserted with the following:
page at ffff80be1fcd6600[3ffff80be1fca5cc00 1ffff809e60e28cc0]
Quoted 9 lines
What do all these fields mean? Or where is this printed? The closest I can find is in
cl_page_header_print()which uses the format string:"page@%p[%d %p %d %d %p]\n",But that only has 5 fields in the [], while the output you provided has 6.
Sorry for the confusion, here is my (long) explanation with source snippets:
The LBUG in full
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0())
page at ffff80be1fcd6600[3 ffff80be1fca5cc0 0 1 ffff809e60e28cc0]
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0()) vvp-page at ffff80be1fcd6650(0:0) vm at ffff7e02fa0e60c0 5000000000001029 4:0 ffff80be1fcd6600 1049548 lru
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0()) lov-page at ffff80be1fcd6690, comp index: 0, gen: 0
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0()) osc-page at ffff80be1fcd66c8 1049548: 1< 0x845fed 2 0 - - > 2< 4298948608 0
4096 0x0 0x420 | (null) ffff809e473708e8 ffff80be199c1e40 > 3< 0
0 0 > 4< 5 13 64 0 + | - - - - > 5< - - - - | 0 - | 0 - ->
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0()) end page at ffff80be1fcd6600
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0()) allowed_transitions[old][state]
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0()) ASSERTION( 0 ) failed:
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0()) LBUG
The source snippet around L342
...
339
340 ENTRY; 341 old = page->cp_state; 342 PASSERT(env, page, allowed_transitions[old][state]); 343 CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old,
state);
344 PASSERT(env, page, page->cp_state == old);
345 PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner
!= NULL));
346
347 cs_pagestate_dec(page->cp_obj, page->cp_state);
348 cs_pagestate_inc(page->cp_obj, state);
The PASSERT macro ...
...
# define PASSERT(env, page, expr) \
do { \
if (unlikely(!(expr))) { \
CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
LASSERT(0); \
} \
} while (0)
#else /* !LIBCFS_DEBUG */
.../**
* Helper macro, dumping detailed information about \a page into a log. */
#define CL_PAGE_DEBUG(mask, env, page, format, ...) \
do { \
if (cfs_cdebug_show(mask, DEBUG_SUBSYSTEM)) { \
LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, NULL); \
cl_page_print(env, &msgdata, lu_cdebug_printer, page); \
CDEBUG(mask, format , ## __VA_ARGS__); \
} \} while (0)
Almost there
...
1062 /**
1063 * Prints human readable representation of \a pg to the \a f.
1064 */
1065 void cl_page_print(const struct lu_env *env, void *cookie,
1066 lu_printer_t printer, const struct cl_page *pg)
1067 {
1068 const struct cl_page_slice *slice;
1069 int result = 0;
1070
1071 cl_page_header_print(env, cookie, printer, pg);
1072 list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
1073 if (slice->cpl_ops->cpo_print != NULL)
1074 result = (*slice->cpl_ops->cpo_print)(env,
slice,
1075
cookie, printer);
Finally the 'good stuff'
...
1051 void cl_page_header_print(const struct lu_env *env, void *cookie,
1052 lu_printer_t printer, const structcl_page *pg)
1053 {
1054 (*printer)(env, cookie,
1055 "page@%p[%d %p %d %d %p]\n", 1056 pg, atomic_read(&pg->cp_ref), pg->cp_obj, 1057 pg->cp_state, pg->cp_type, 1058 pg->cp_owner);
1059 }
...
Given the above along with the first line from the LBUG ....
LustreError: 86407:0:(cl_page.c:342:cl_page_state_set0()) page at ffff80be1fcd6600[3 ffff80be1fca5cc0 0 1 ffff809e60e28cc0]
Implies the following:
pg == ffff80be1fcd6600pg->cp_ref == 3pg->cp_obj == ffff80be1fca5cc0pg->cp_state == 0 (state == CPS_CACHED)pg->cp_type == 1pg->cp_owner == ffff809e60e28cc0
Also known caller was cl_page_assume+0xdc/0x3e0 [obdclass] (from the backtrace)
So we know the on-stack / register value of state == CPS_OWNED
The backtrace:
...
dump_backtrace+0x0/0x248 show_stack+0x24/0x30 dump_stack+0xbc/0xf4 libcfs_call_trace+0xec/0x120 [libcfs] lbug_with_loc+0x4c/0xb8 [libcfs] cl_page_state_set0+0x2b4/0x6e0 [obdclass] cl_page_assume+0xdc/0x3e0 [obdclass] ll_io_read_page+0x144c/0x1de0 [lustre] ll_write_begin+0x3d4/0xee8 [lustre]
and the contents of cl_page_assume
596 void cl_page_assume(const struct lu_env *env,
597 struct cl_io *io, struct cl_page *pg)
598 {
599 const struct cl_page_slice *slice;
600
601 PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
602
603 ENTRY;
604 io = cl_io_top(io);
605
606 list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
607 if (slice->cpl_ops->cpo_assume != NULL)
608 (*slice->cpl_ops->cpo_assume)(env, slice,
io);
609 }
610
611 PASSERT(env, pg, pg->cp_owner == NULL); 612 pg->cp_owner = cl_io_top(io); 613 cl_page_owner_set(pg); 614 cl_page_state_set(env, pg, CPS_OWNED); 615 EXIT;
616 }
While I believe my conclusion to be correct I am certainly open to being dead wrong.
Quoted 36 lines · NeilBrown
Quoted 10 lines · Shaun Tancheff
Ensure all uses of
cl_page->cp_stateare smp-safe.LBUG Output:
In__cl_page_state_set()
..
old =page->cp_state;
PASSERT(env, page, allowed_transitions[old][state]);
..
Asserted with the following:
page at ffff80be1fcd6600[3ffff80be1fca5cc00 1ffff809e60e28cc0]What do all these fields mean? Or where is this printed? The closest I can find is in
cl_page_header_print()which uses the format string:"page@%p[%d %p %d %d %p]\n",But that only has 5 fields in the [], while the output you provided has 6.
Quoted 4 lines
However
cp_state0 (CPS_CACHED) to 1 (CPS_OWNED) is a valid transition leading to the conclusion thatcp_statebecame 0 during the assertion.Where is the evidence that this was the transition that was happening? If it was in some part of the LBUG output that wasn't quoted - then please fix that by quoting the entire LBUG output.
We need to understand which change happened at this time to cause the race.
Then we need to explain why the added barriers actually close the race window.
Quoted 9 lines
If these barrier do actually fix a race, then there is something *very* wrong.
The comments say that changes tocp_stateare protected by the page lock on the corresponding VM page.
Locking and unlocking a VM page entails sufficient barriers that changes made while one thread holds the lock will be visible to another thread once it also gets the lock.
True, but a page lock is just a bit on the page flags. And we absolutely wait on I/O while the
page is 'locked' so a ping-pong (our task *can* sleep) is entirely reasonable, as far as
I understand anyway. Since there is multiple state-changes during a single page lock/unlock
there is a race, however rare and unlikely, and it's arch specific based on what I understand
of the memory barrier rules.
So the only possible explanation for a race as you suspect, is if the
Quoted 285 lines
vmpage *isn't* locked when something changes
cp_state, and that would be bad.Thanks,
NeilBrownQuoted 11 lines
Signed-off-by Shaun Tancheff <[email protected]>--- fs/lustre/include/cl_object.h | 11 +++++++++++ fs/lustre/llite/rw26.c | 2 +- fs/lustre/llite/vvp_page.c | 6 +++--- fs/lustre/obdclass/cl_page.c | 34 +++++++++++++++++++-------------- fs/lustre/obdecho/echo_client.c | 2 +- fs/lustre/osc/osc_cache.c | 18 +++++++++-------- 6 files changed, 46 insertions(+), 27 deletions(-)Review
Quoted 1 line
(quoted diff) +0
-0
b/fs/lustre/include/cl_object.hfs/lustre/include/cl_object.h struct cl_page { · line 755
Quoted 8 lines
index
691c2f5da53a..d6e1f6f05f50100644fs/lustre/include/cl_object.h +1
-0@@ -752,6 +752,17 @@ struct cl_page {struct cl_sync_io *cp_sync_io; }; +static inline enum cl_page_state cl_page_state_get(const struct cl_page*pg)
fs/lustre/llite/rw26.c static ssize_t ll_direct_IO_seg(const struct lu_env
Quoted 18 lines
fs/lustre/include/cl_object.h +10
-0+{ +/* +* Paired with smp_store_release in cl_page_state_set_trust +* and ensures that we see the most recent value of cp_state +* even when the last modification was not performed on the +* current processor +*/ +return smp_load_acquire (&pg->cp_state); +} +/** * Per-layer part of cl_page. * fs/lustre/llite/rw26.c +0
-0@@ -200,7 +200,7 @@ static ssize_t ll_direct_IO_seg(const struct lu_env*env, struct
cl_io*io,fs/lustre/llite/vvp_page.c static int vvp_page_make_ready(const struct lu_env
Quoted 13 lines
fs/lustre/llite/rw26.c +1
-1rc =cl_page_own (env, io, clp);if (rc) { -LASSERT (clp->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (clp) ==CPS_FREEING );cl_page_put (env, clp);break ; }fs/lustre/llite/vvp_page.c +0
-0@@ -323,18 +323,18 @@ static int vvp_page_make_ready(const struct lu_env*env,
Quoted 6 lines
fs/lustre/llite/vvp_page.c +1
-1lock_page (vmpage);if (clear_page_dirty_for_io (vmpage)) { -LASSERT (pg->cp_state ==CPS_CACHED ); +LASSERT (cl_page_state_get (pg) ==CPS_CACHED );/* This actually clears the dirty bit in the radix tree. */ fs/lustre/llite/vvp_page.c static int vvp_page_make_ready(const struct lu_env *env,
Quoted 10 lines
fs/lustre/llite/vvp_page.c +1
-1set_page_writeback (vmpage);CL_PAGE_HEADER (D_PAGE , env, pg,"readied\n" ); - }else if (pg->cp_state ==CPS_PAGEOUT ) { + }else if (cl_page_state_get (pg) ==CPS_PAGEOUT ) {/* is it possible for osc_flush_async_page() to already * make it ready? */ result = -EALREADY ; }else { CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state%d.\n",
fs/lustre/obdclass/cl_page.c static void cl_page_free(const struct lu_env *env,
Quoted 10 lines
fs/lustre/llite/vvp_page.c +1
-1- pg->cp_state); + cl_page_state_get(pg));LBUG (); }unlock_page (vmpage);fs/lustre/obdclass/cl_page.c +0
-0@@ -97,7 +97,7 @@ static void cl_page_free(const struct lu_env *env,struct
cl_page*page)fs/lustre/obdclass/cl_page.c static void cl_page_free(const struct lu_env *env,
Quoted 9 lines
fs/lustre/obdclass/cl_page.c +1
-1PASSERT (env, page,list_empty (&page->cp_batch));PASSERT (env, page, !page->cp_owner); -PASSERT (env, page, page->cp_state ==CPS_FREEING ); +PASSERT (env, page,cl_page_state_get (page) ==CPS_FREEING ); while ((slice = list_first_entry_or_null(&page->cp_layers, struct cl_page_slice,@@ -119,8 +119,14 @@ static void cl_page_free(const struct lu_env *env,struct
cl_page*page)fs/lustre/obdclass/cl_page.c static void __cl_page_state_set(const struct
Quoted 17 lines
fs/lustre/obdclass/cl_page.c +8
-2static inline void cl_page_state_set_trust (struct cl_page *page,enum cl_page_state state) { -/* bypass const. */ - *(enum cl_page_state *)&page->cp_state = state; +/* +* Paired with smp_load_acquire in cl_page_state_get +* and ensures that we see the most recent value of cp_state +* is available even when the next access is not performed on the +* current processor. +* Note we also cast away const as the only modifier of cp_state. +*/ +smp_store_release ((enum cl_page_state *)&page->cp_state, state); } struct cl_page *cl_page_alloc(const struct lu_env *env,@@ -270,10 +276,10 @@ static void __cl_page_state_set(const struct
lu_env*env,fs/lustre/obdclass/cl_page.c void cl_page_put(const struct lu_env *env, struct
Quoted 13 lines
fs/lustre/obdclass/cl_page.c +2
-2} }; - old = page->cp_state; + old =cl_page_state_get (page);PASSERT (env, page, allowed_transitions[old][state]);CL_PAGE_HEADER (D_TRACE , env, page,"%d -> %d\n" , old, state); -PASSERT (env, page, page->cp_state == old); +PASSERT (env, page,cl_page_state_get (page) == old);PASSERT (env, page,equi (state ==CPS_OWNED , page->cp_owner));cl_page_state_set_trust (page, state); }@@ -313,7 +319,7 @@ void cl_page_put(const struct lu_env *env, struct
cl_page*page)fs/lustre/obdclass/cl_page.c int cl_page_is_owned(const struct cl_page *pg, const
Quoted 18 lines
fs/lustre/obdclass/cl_page.c +2
-2refcount_read(&page->cp_ref));if (refcount_dec_and_test (&page->cp_ref)) { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING );LASSERT (refcount_read (&page->cp_ref) ==0 );PASSERT (env, page, !page->cp_owner);@@ -378,7 +384,7 @@ void __cl_page_disown(const struct lu_env *env,const struct cl_page_slice *slice;enum cl_page_state state; - state = pg->cp_state; + state =cl_page_state_get (pg);cl_page_owner_clear (pg);if (state ==CPS_OWNED )@@ -402,7 +408,7 @@ int cl_page_is_owned(const struct cl_page *pg, conststruct
cl_io*io)fs/lustre/obdclass/cl_page.c static int __cl_page_own(const struct lu_env *env,
Quoted 9 lines
fs/lustre/obdclass/cl_page.c +1
-1struct cl_io *top =cl_io_top ((struct cl_io *)io);LINVRNT (cl_object_same (pg->cp_obj, io->ci_obj)); -return pg->cp_state ==CPS_OWNED && pg->cp_owner == top; +return cl_page_state_get (pg) ==CPS_OWNED && pg->cp_owner == top; }EXPORT_SYMBOL (cl_page_is_owned);@@ -434,7 +440,7 @@ static int __cl_page_own(const struct lu_env *env,struct
cl_io*io,fs/lustre/obdclass/cl_page.c static int __cl_page_own(const struct lu_env *env,
Quoted 9 lines
fs/lustre/obdclass/cl_page.c +1
-1io =cl_io_top (io); -if (pg->cp_state ==CPS_FREEING ) { +if (cl_page_state_get (pg) ==CPS_FREEING ) { result = -ENOENT ; goto out; }@@ -453,7 +459,7 @@ static int __cl_page_own(const struct lu_env *env,struct
cl_io*io,fs/lustre/obdclass/cl_page.c static void __cl_page_delete(const struct lu_env
Quoted 9 lines
fs/lustre/obdclass/cl_page.c +1
-1PASSERT (env, pg, !pg->cp_owner); pg->cp_owner =cl_io_top (io);cl_page_owner_set (pg); -if (pg->cp_state !=CPS_FREEING ) { +if (cl_page_state_get (pg) !=CPS_FREEING ) {cl_page_state_set (env, pg,CPS_OWNED ); }else {__cl_page_disown (env, io, pg);@@ -593,7 +599,7 @@ static void __cl_page_delete(const struct lu_env*env, struct
cl_page*pg)fs/lustre/obdclass/cl_page.c int cl_page_make_ready(const struct lu_env *env,
Quoted 18 lines
fs/lustre/obdclass/cl_page.c +2
-2{const struct cl_page_slice *slice; -PASSERT (env, pg, pg->cp_state !=CPS_FREEING ); +PASSERT (env, pg,cl_page_state_get (pg) !=CPS_FREEING );/* * Sever all ways to obtain new pointers to @pg. @@ -756,7 +762,7 @@ void cl_page_completion(const struct lu_env *env,const struct cl_page_slice *slice;PASSERT (env, pg, crt <CRT_NR ); -PASSERT (env, pg, pg->cp_state ==cl_req_type_state (crt)); +PASSERT (env, pg,cl_page_state_get (pg) ==cl_req_type_state (crt));CL_PAGE_HEADER (D_TRACE , env, pg,"%d %d\n" , crt, ioret);@@ -805,7 +811,7 @@ int cl_page_make_ready(const struct lu_env *env,struct
cl_page*pg,fs/lustre/obdclass/cl_page.c void cl_page_header_print(const struct lu_env *env,
Quoted 9 lines
fs/lustre/obdclass/cl_page.c +1
-1}if (result >=0 ) { -PASSERT (env, pg, pg->cp_state ==CPS_CACHED ); +PASSERT (env, pg,cl_page_state_get (pg) ==CPS_CACHED );cl_page_io_start (env, pg, crt); result =0 ; }@@ -870,7 +876,7 @@ void cl_page_header_print(const struct lu_env *env,void *cookie,
Review
Quoted 9 lines
fs/lustre/obdclass/cl_page.c +1
-1(*printer)(env, cookie,"page@%p[%d %p %d %d %p]\n" , pg,refcount_read (&pg->cp_ref), pg->cp_obj, - pg->cp_state, pg->cp_type, +cl_page_state_get (pg), pg->cp_type, pg->cp_owner); }EXPORT_SYMBOL (cl_page_header_print);(quoted diff) +0
-0
b/fs/lustre/obdecho/echo_client.cfs/lustre/obdecho/echo_client.c static int cl_echo_object_brw(struct echo_object
Quoted 4 lines
index
317123fd27cb..d879f109e641100644fs/lustre/obdecho/echo_client.c +0
-0@@ -1046,7 +1046,7 @@ static int cl_echo_object_brw(struct echo_object*eco, int rw, u64 offset,
fs/lustre/osc/osc_cache.c static int osc_extent_truncate(struct osc_extent
Quoted 13 lines
fs/lustre/obdecho/echo_client.c +1
-1rc =cl_page_own (env, io, clp);if (rc) { -LASSERT (clp->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (clp) ==CPS_FREEING );cl_page_put (env, clp);break ; }fs/lustre/osc/osc_cache.c +0
-0@@ -1045,7 +1045,7 @@ static int osc_extent_truncate(struct osc_extent*ext,
pgoff_ttrunc_index,fs/lustre/osc/osc_cache.c static int osc_completion(const struct lu_env
Quoted 9 lines
fs/lustre/osc/osc_cache.c +1
-1cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING );LASSERT (0 ); }@@ -1356,10 +1356,12 @@ static int osc_completion(const struct lu_env*env, struct
osc_async_page*oap,fs/lustre/osc/osc_cache.c bool osc_page_gang_lookup(const struct lu_env
Quoted 17 lines
fs/lustre/osc/osc_cache.c +6
-4int srvlock; cmd &= ~OBD_BRW_NOQUOTA ; -LASSERTF (equi (page->cp_state ==CPS_PAGEIN , cmd ==OBD_BRW_READ ), -"cp_state:%u, cmd:%d\n" , page->cp_state, cmd); -LASSERTF (equi (page->cp_state ==CPS_PAGEOUT , cmd ==OBD_BRW_WRITE ), -"cp_state:%u, cmd:%d\n" , page->cp_state, cmd); +LASSERTF (equi (cl_page_state_get (page) ==CPS_PAGEIN , + cmd ==OBD_BRW_READ ), +"cp_state:%u, cmd:%d\n" ,cl_page_state_get (page), cmd); +LASSERTF (equi (cl_page_state_get (page) ==CPS_PAGEOUT , + cmd ==OBD_BRW_WRITE ), +"cp_state:%u, cmd:%d\n" ,cl_page_state_get (page), cmd);LASSERT (opg->ops_transfer_pinned); crt = cmd ==OBD_BRW_READ ?CRT_READ :CRT_WRITE ;@@ -3061,7 +3063,7 @@ bool osc_page_gang_lookup(const struct lu_env*env, struct
cl_io*io,fs/lustre/osc/osc_cache.c static bool check_and_discard_cb(const struct
Quoted 9 lines
fs/lustre/osc/osc_cache.c +1
-1page = ops->ops_cl.cpl_page;LASSERT (page->cp_type ==CPT_CACHEABLE ); -if (page->cp_state ==CPS_FREEING ) +if (cl_page_state_get (page) ==CPS_FREEING )continue ;cl_page_get (page);@@ -3142,7 +3144,7 @@ static bool check_and_discard_cb(const struct
lu_env*env, structcl_io*io,Quoted 8 lines
fs/lustre/osc/osc_cache.c +1
-1cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING ); } }
Quoted 13 lines
fs/lustre/osc/osc_cache.c static bool discard_cb(const struct lu_env *env,
Quoted 1 line
fs/lustre/osc/osc_cache.c +0
-0@@ -3169,7 +3171,7 @@ static bool discard_cb(const struct lu_env *env,struct
cl_io*io,Quoted 10 lines
fs/lustre/osc/osc_cache.c +1
-2cl_page_discard (env, io, page);cl_page_disown (env, io, page); }else { -LASSERT (page->cp_state ==CPS_FREEING ); +LASSERT (cl_page_state_get (page) ==CPS_FREEING ); }return true; --2.17.1
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20190606/6bdc81e1/attachment-0001.html>
PATCH 1/1 Place a memory barrier around cp_state changes
Quoted 6 lines · Shaun Tancheff
Quoted 5 lines · NeilBrown
Quoted 2 lines · Shaun Tancheff
Ensure all uses of
cl_page->cp_stateare smp-safe.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
LBUG Output:
In__cl_page_state_set()
..
old =page->cp_state;
PASSERT(env, page, allowed_transitions[old][state]);
..
Asserted with the following:
page at ffff80be1fcd6600[3ffff80be1fca5cc00 1ffff809e60e28cc0]
I now see that this space..... ^^^
shouldn't be there..
.....
Quoted 6 lines
Also known caller was
cl_page_assume+0xdc/0x3e0[obdclass] (from the backtrace)
So we know the on-stack / register value of state ==CPS_OWNEDThe backtrace:
...
Quoted 9 lines
dump_backtrace+0x0/0x248 show_stack+0x24/0x30 dump_stack+0xbc/0xf4 libcfs_call_trace+0xec/0x120 [libcfs] lbug_with_loc+0x4c/0xb8 [libcfs] cl_page_state_set0+0x2b4/0x6e0 [obdclass] cl_page_assume+0xdc/0x3e0 [obdclass] ll_io_read_page+0x144c/0x1de0 [lustre] ll_write_begin+0x3d4/0xee8 [lustre]
Thanks for this extra detail.
I agree that the evidence suggests that page->cp_state changed (to CPS_CACHED) during the processing of the PASSERT().
It isn't clear to me though that this problem can be fixed just by adding barriers - particularly if this was on x86_64 as I suspect it is. Intel architecture processors have strongly ordered memory semantics. I'm not entirely sure what that means, but as smp_load_acquire() is very nearly a no-op on x86, I doubt that the memory barriers you added actually fix anything there.
My understanding (based on a fairly shallow reading of the code) is that a partial-page write is being attempted, and the target page was not in memory, so it was read in synchronously.
The fragment of ll_io_read_page happening is:
if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
rc = cl_sync_io_wait(env, anchor, 0);
cl_page_assume(env, io, page);So cl_sync_io_wait() has just run. The page was (presumably) previously owned by something else. cl_sync_io_wait() waited for csi_sync_nr to drop to zero, which should say it isn't owned any more. cl_page_assume() then takes ownership.
But it seems that the old ownership wasn't quite gone yet.
The VM page remains locked the whole time, so the barriers implicit in locking and unlocking a page do not some into play here. There is an implicit barrier on the 'store' side when cl_sync_io_note() calls atomic_dec_and_test() to decrement csi_sync_nr, but I cannot see any implicit barrier on the 'load' side. If l_wait_event() in cl_sync_io_wait() took a spinlock, that would be enough. But it is possible for the wait to complete without any locking.
So I can just about see the possibility of the race that you think you have hit. But it couldn't happen on x86, as it has strong memory ordering.
What CPU architecture did this ASSERT trigger on?
Thanks,
NeilBrown
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20190611/656d36a4/attachment.sig>
PATCH 1/1 Place a memory barrier around cp_state changes
Quoted 41 lines · NeilBrown
Quoted 6 lines · Shaun Tancheff
Quoted 5 lines · NeilBrown
Quoted 2 lines · Shaun Tancheff
Ensure all uses of
cl_page->cp_stateare smp-safe.Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
LBUG Output:
In__cl_page_state_set()
..
old =page->cp_state;
PASSERT(env, page, allowed_transitions[old][state]);
..
Asserted with the following:
page at ffff80be1fcd6600[3ffff80be1fca5cc00 1ffff809e60e28cc0]I now see that this space..... ^^^
shouldn't be there..
.....Quoted 6 lines
Also known caller was
cl_page_assume+0xdc/0x3e0[obdclass] (from the backtrace)
So we know the on-stack / register value of state ==CPS_OWNEDThe backtrace:
...Quoted 9 lines
dump_backtrace+0x0/0x248 show_stack+0x24/0x30 dump_stack+0xbc/0xf4 libcfs_call_trace+0xec/0x120 [libcfs] lbug_with_loc+0x4c/0xb8 [libcfs] cl_page_state_set0+0x2b4/0x6e0 [obdclass] cl_page_assume+0xdc/0x3e0 [obdclass] ll_io_read_page+0x144c/0x1de0 [lustre] ll_write_begin+0x3d4/0xee8 [lustre]Thanks for this extra detail.
I agree that the evidence suggests thatpage->cp_statechanged (toCPS_CACHED) during the processing of thePASSERT().
Quoted 7 lines
It isn't clear to me though that this problem can be fixed just by adding barriers - particularly if this was on
x86_64as I suspect it is. Intel architecture processors have strongly ordered memory semantics. I'm not entirely sure what that means, but assmp_load_acquire()is very nearly a no-op on x86, I doubt that the memory barriers you added actually fix anything there.
Agree. This occurred on aarch64. I suspect one of the reasons that this has not
been seen is that the majority of users are x86_64. It appears to be very rare on
aarch64 but it could also, theoretically, happen on the other relaxed architectures.
Quoted 25 lines
My understanding (based on a fairly shallow reading of the code) is that a partial-page write is being attempted, and the target page was not in memory, so it was read in synchronously.
The fragment ofll_io_read_pagehappening is:if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */ rc = cl_sync_io_wait(env, anchor, 0); cl_page_assume(env, io, page);So
cl_sync_io_wait()has just run. The page was (presumably) previously owned by something else.cl_sync_io_wait()waited forcsi_sync_nrto drop to zero, which should say it isn't owned any more.cl_page_assume()then takes ownership.
But it seems that the old ownership wasn't quite gone yet.The VM page remains locked the whole time, so the barriers implicit in locking and unlocking a page do not some into play here. There is an implicit barrier on the 'store' side when
cl_sync_io_note()callsatomic_dec_and_test()to decrementcsi_sync_nr, but I cannot see any implicit barrier on the 'load' side. Ifl_wait_event()incl_sync_io_wait()took a spinlock, that would be enough. But it is possible for the wait to complete without any locking.
Quoted 6 lines
So I can just about see the possibility of the race that you think you have hit. But it couldn't happen on x86, as it has strong memory ordering.
What CPU architecture did this ASSERT trigger on?
Agree with everything above, unfortunately was aarch64.
Quoted 3 lines
Thanks,
NeilBrown
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20190610/b613803d/attachment.html>
PATCH 1/1 Place a memory barrier around cp_state changes
Quoted 11 lines · Shaun Tancheff
Quoted 7 lines · NeilBrown
Quoted 1 line
So I can just about see the possibility of the race that you think you have hit. But it couldn't happen on x86, as it has strong memory ordering.
What CPU architecture did this ASSERT trigger on?
Agree with everything above, unfortunately was aarch64.
Ahh.
OK, I'm now reasonably convinced that the patch probably fixes the bug. It is a shame that there is no standard barrier between the wake_up and the wait_event(), but there isn't and we need to work with that.
But I still don't like the patch.... :-(
I have a few problems with it.
Firstly, I generally don't like accessor functions. They tend to hide useful information.
Rather than cl_page_state_get() I would rather have
smp_load_acquire(&pg->cp_state)
where ever it is needed, with a comment explaining why it is needed in that particular context.
Some people seem to like accessor functions - there certainly are a lot of them in the kernel - but I don't think they help.
Having an accessor function that just adds a barrier is, I think, particularly ugly as the comment can only be general in nature, and so doesn't really help the reader to understand why the barrier is needed. But even that isn't completely without precedent. "key_read_state()" in include/linux/key.h simply wraps smp_load_acquire(), and the comment there doesn't really tell me anything useful.
"inet_sk_state_load" is slightly better it says it is for use in places where the socket lock isn't held - and names some of those.
My second problem is that has "get" in the accessor function name is potentially misleading and "get" normally has a matching "put", and usually increments a refcount. Having "read" or "load" in place of "get" in the accessor function name would remove my second objection.
My third objection is that the bug doesn't exist in the upstream client code. I'm not sure whether this is luck or good management, but either way it is gone.
cl_sync_io_wait() has been changed to always take csi_waitq.lock. This avoids the need for anchor->csi_barrier, and provides a more robust way to ensure that cl_sync_io_note() is completely finished before cl_sync_io_wait() completes.
That isn't quite the outcome I was expecting...
I'm sorry if I seem to have been rather harsh on your patch - my intention was only to make sure it was the best it could be.
It really needed to have the stack trace and the arch were the LBUG happened to be at all complete.
I would have been nice if it identified the particular code where the barrier was needed: between cl_sync_io_note() and cl_sync_io_wait() because that mediates and ownership transition that happens while the VM page is locked.
The rest are probably just issues of taste.
Thanks,
NeilBrown
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20190612/6b656056/attachment.sig>
PATCH 1/1 Place a memory barrier around cp_state changes
Quoted 28 lines · NeilBrown
Quoted 11 lines · Shaun Tancheff
Quoted 7 lines · NeilBrown
Quoted 1 line
So I can just about see the possibility of the race that you think you have hit. But it couldn't happen on x86, as it has strong memory ordering.
What CPU architecture did this ASSERT trigger on?
Agree with everything above, unfortunately was aarch64.
Ahh.
OK, I'm now reasonably convinced that the patch probably fixes the bug. It is a shame that there is no standard barrier between thewake_upand thewait_event(), but there isn't and we need to work with that.But I still don't like the patch.... :-(
I have a few problems with it.
Firstly, I generally don't like accessor functions. They tend to hide useful information.
Rather thancl_page_state_get()I would rather havesmp_load_acquire(&pg->cp_state)
Quoted 14 lines
where ever it is needed, with a comment explaining why it is needed in that particular context.
Some people seem to like accessor functions - there certainly are a lot of them in the kernel - but I don't think they help.
Having an accessor function that just adds a barrier is, I think, particularly ugly as the comment can only be general in nature, and so doesn't really help the reader to understand why the barrier is needed. But even that isn't completely without precedent. "key_read_state()" ininclude/linux/key.hsimply wrapssmp_load_acquire(), and the comment there doesn't really tell me anything useful.
"inet_sk_state_load" is slightly better it says it is for use in places where the socket lock isn't held - and names some of those.
Understood and a fair assessment.
Quoted 5 lines
My second problem is that has "get" in the accessor function name is potentially misleading and "get" normally has a matching "put", and usually increments a refcount. Having "read" or "load" in place of "get" in the accessor function name would remove my second objection.
Yeah, I was trying to balance with the set .. but really I'm not enamored with that either :P
Quoted 4 lines
My third objection is that the bug doesn't exist in the upstream client code. I'm not sure whether this is luck or good management, but either way it is gone.
This is really good to know. Personally I was itching to refactor the whole lot but this gives a me a much better path.
Quoted 17 lines
cl_sync_io_wait()has been changed to always take csi_waitq.lock. This avoids the need foranchor->csi_barrier, and provides a more robust way to ensure thatcl_sync_io_note()is completely finished beforecl_sync_io_wait()completes.That isn't quite the outcome I was expecting...
I'm sorry if I seem to have been rather harsh on your patch - my intention was only to make sure it was the best it could be.
It really needed to have the stack trace and the arch were the LBUG happened to be at all complete.
I would have been nice if it identified the particular code where the barrier was needed: betweencl_sync_io_note()andcl_sync_io_wait()because that mediates and ownership transition that happens while the VM page is locked.
Quoted 5 lines
The rest are probably just issues of taste.
Thanks,
NeilBrown
Thanks I really appreciate the time need to consider the issue. And I agree any patch needs to be as good and tasteful as we can make them.
Regards,
Shaun
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20190611/7dffcc07/attachment.html>
PATCH 1/1 Place a memory barrier around cp_state changes
As an interested observer (thanks to both of you for an interesting exchange), Shaun, what?s your plan going forward for the OpenSFS/WC branch? It sounds like you?re thinking you?ll try to emulate what Neil did upstream? (which sounds good to me, I always prefer avoiding explicit memory barriers if reasonable) Will you be opening an LU for this one?
Thanks,
- Patrick
________________________________
From: lustre-devel <[email protected]> on behalf of Shaun Tancheff <[email protected]>
Sent: Tuesday, June 11, 2019 11:36:13 PM
To: NeilBrown
Cc: Bobi Jam; lustre-devel at lists.lustre.org
Subject: Re: [lustre-devel] [PATCH] Place a memory barrier around cp_state changes
On Tue, Jun 11, 2019 at 11:09 PM NeilBrown <neilb at suse.com<mailto:[email protected]>> wrote:
Quoted 11 lines · Shaun Tancheff
Quoted 7 lines · On Mon, Jun 10, 2019 at 7:49 PM NeilBrown <neilb at…
Quoted 1 line
So I can just about see the possibility of the race that you think you have hit. But it couldn't happen on x86, as it has strong memory ordering.
What CPU architecture did this ASSERT trigger on?
Agree with everything above, unfortunately was aarch64.
Ahh.
OK, I'm now reasonably convinced that the patch probably fixes the bug. It is a shame that there is no standard barrier between the wake_up and the wait_event(), but there isn't and we need to work with that.
But I still don't like the patch.... :-(
I have a few problems with it.
Firstly, I generally don't like accessor functions. They tend to hide useful information.
Rather than cl_page_state_get() I would rather have
smp_load_acquire(&pg->cp_state)
where ever it is needed, with a comment explaining why it is needed in that particular context.
Some people seem to like accessor functions - there certainly are a lot of them in the kernel - but I don't think they help.
Having an accessor function that just adds a barrier is, I think, particularly ugly as the comment can only be general in nature, and so doesn't really help the reader to understand why the barrier is needed. But even that isn't completely without precedent. "key_read_state()" in include/linux/key.h simply wraps smp_load_acquire(), and the comment there doesn't really tell me anything useful.
"inet_sk_state_load" is slightly better it says it is for use in places where the socket lock isn't held - and names some of those.
Understood and a fair assessment.
My second problem is that has "get" in the accessor function name is potentially misleading and "get" normally has a matching "put", and usually increments a refcount. Having "read" or "load" in place of "get" in the accessor function name would remove my second objection.
Yeah, I was trying to balance with the set .. but really I'm not enamored with that either :P
My third objection is that the bug doesn't exist in the upstream client code. I'm not sure whether this is luck or good management, but either way it is gone.
This is really good to know. Personally I was itching to refactor the whole lot but this gives a me a much better path.
cl_sync_io_wait() has been changed to always take csi_waitq.lock. This avoids the need for anchor->csi_barrier, and provides a more robust way to ensure that cl_sync_io_note() is completely finished before cl_sync_io_wait() completes.
That isn't quite the outcome I was expecting...
I'm sorry if I seem to have been rather harsh on your patch - my intention was only to make sure it was the best it could be.
It really needed to have the stack trace and the arch were the LBUG happened to be at all complete.
I would have been nice if it identified the particular code where the barrier was needed: between cl_sync_io_note() and cl_sync_io_wait() because that mediates and ownership transition that happens while the VM page is locked.
The rest are probably just issues of taste.
Thanks,
NeilBrown
Thanks I really appreciate the time need to consider the issue. And I agree any patch needs to be as good and tasteful as we can make them.
Regards,
Shaun
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20190612/b72ea4cf/attachment-0001.html>
PATCH 1/1 Place a memory barrier around cp_state changes
Quoted 6 lines · Patrick Farrell
As an interested observer (thanks to both of you for an interesting exchange), Shaun, what?s your plan going forward for the OpenSFS/WC branch? It sounds like you?re thinking you?ll try to emulate what Neil did upstream? (which sounds good to me, I always prefer avoiding explicit memory barriers if reasonable) Will you be opening an LU for this one?
Yes, I will be opening an LU for this. My plan is to emulate upstream.
Quoted 101 lines
Thanks,
- Patrick
------------------------------
*From:* lustre-devel <[email protected]> on behalf of Shaun Tancheff <[email protected]>
*Sent:* Tuesday, June 11, 2019 11:36:13 PM
*To:* NeilBrown
*Cc:* Bobi Jam; lustre-devel at lists.lustre.org
*Subject:* Re: [lustre-devel] [PATCH] Place a memory barrier aroundcp_statechangesOn Tue, Jun 11, 2019 at 11:09 PM NeilBrown <[email protected]> wrote:
Quoted 11 lines · Shaun Tancheff
Quoted 7 lines · NeilBrown
Quoted 1 line
So I can just about see the possibility of the race that you think you have hit. But it couldn't happen on x86, as it has strong memory ordering.
What CPU architecture did this ASSERT trigger on?
Agree with everything above, unfortunately was aarch64.
Ahh.
OK, I'm now reasonably convinced that the patch probably fixes the bug. It is a shame that there is no standard barrier between thewake_upand thewait_event(), but there isn't and we need to work with that.But I still don't like the patch.... :-(
I have a few problems with it.
Firstly, I generally don't like accessor functions. They tend to hide useful information.
Rather thancl_page_state_get()I would rather havesmp_load_acquire(&pg->cp_state)where ever it is needed, with a comment explaining why it is needed in that particular context.
Some people seem to like accessor functions - there certainly are a lot of them in the kernel - but I don't think they help.
Having an accessor function that just adds a barrier is, I think, particularly ugly as the comment can only be general in nature, and so doesn't really help the reader to understand why the barrier is needed. But even that isn't completely without precedent. "key_read_state()" ininclude/linux/key.hsimply wrapssmp_load_acquire(), and the comment there doesn't really tell me anything useful.
"inet_sk_state_load" is slightly better it says it is for use in places where the socket lock isn't held - and names some of those.Understood and a fair assessment.
My second problem is that has "get" in the accessor function name is potentially misleading and "get" normally has a matching "put", and usually increments a refcount. Having "read" or "load" in place of "get" in the accessor function name would remove my second objection.
Yeah, I was trying to balance with the set .. but really I'm not enamored with that either :P
My third objection is that the bug doesn't exist in the upstream client code. I'm not sure whether this is luck or good management, but either way it is gone.
This is really good to know. Personally I was itching to refactor the whole lot but this gives a me a much better path.
cl_sync_io_wait()has been changed to always take csi_waitq.lock. This avoids the need foranchor->csi_barrier, and provides a more robust way to ensure thatcl_sync_io_note()is completely finished beforecl_sync_io_wait()completes.That isn't quite the outcome I was expecting...
I'm sorry if I seem to have been rather harsh on your patch - my intention was only to make sure it was the best it could be.
It really needed to have the stack trace and the arch were the LBUG happened to be at all complete.
I would have been nice if it identified the particular code where the barrier was needed: betweencl_sync_io_note()andcl_sync_io_wait()because that mediates and ownership transition that happens while the VM page is locked.The rest are probably just issues of taste.
Thanks,
NeilBrown
Quoted 7 lines
Thanks I really appreciate the time need to consider the issue. And I agree any patch needs to be as good and tasteful as we can make them.
Regards,
Shaun
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20190612/74f85baf/attachment.html>