| Message ID | 20181121101839.30784-17-mwilck@suse.com (mailing list archive) |
|---|---|
| State | Not Applicable, archived |
| Delegated to: | christophe varoqui |
| Headers | show |
| Series | multipath-tools: improve logging at -v3 | expand |
On Wed, Nov 21, 2018 at 11:18:36AM +0100, Martin Wilck wrote: > When there are paths with the same WWID but different sizes, and > coalesce_paths() walks the pathvec, it checks paths _after_ > the current one for size mismatch and sets ACT_REJECT. However, > these paths will be reached in the main loop later, and this time > the already handled paths will not be checked for size mismatch; > thus a map could be created, possibly even with mismatching > devices. > > Fix that by tracking which paths were already discarded, and > skipping them in the main loop later. Previously, multipath would retry if coalesce_paths returned DOMAP_RETRY. With this patch, DOMAP_RETRY isn't returned, so that no longer happens. Is this intentional? Also, I would prefer if coalesce_paths always used named constants for the return values, instead of converting them to numbers. -Ben > Signed-off-by: Martin Wilck <mwilck@suse.com> > --- > libmultipath/configure.c | 33 +++++++++++++++++++++++++-------- > libmultipath/util.h | 16 ++++++++++++++++ > 2 files changed, 41 insertions(+), 8 deletions(-) > > diff --git a/libmultipath/configure.c b/libmultipath/configure.c > index 406cd4c9..4ed3cc23 100644 > --- a/libmultipath/configure.c > +++ b/libmultipath/configure.c > @@ -1009,6 +1009,7 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, > vector pathvec = vecs->pathvec; > struct config *conf; > int allow_queueing; > + uint64_t *size_mismatch_seen; > > /* ignore refwwid if it's empty */ > if (refwwid && !strlen(refwwid)) > @@ -1019,6 +1020,14 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, > pp1->mpp = NULL; > } > } > + > + if (VECTOR_SIZE(pathvec) == 0) > + return 0; > + size_mismatch_seen = calloc((VECTOR_SIZE(pathvec) - 1) / 64 + 1, > + sizeof(uint64_t)); > + if (size_mismatch_seen == NULL) > + return 1; > + > vector_foreach_slot (pathvec, pp1, k) { > int invalid; > /* skip this path for some reason */ > @@ -1038,8 +1047,8 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, > continue; > } > > - /* 2. if path already coalesced */ > - if (pp1->mpp) > + /* 2. if path already coalesced, or seen and discarded */ > + if (pp1->mpp || is_bit_set_in_array(k, size_mismatch_seen)) > continue; > > /* 3. if path has disappeared */ > @@ -1088,9 +1097,10 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, > * ouch, avoid feeding that to the DM > */ > condlog(0, "%s: size %llu, expected %llu. " > - "Discard", pp2->dev_t, pp2->size, > + "Discard", pp2->dev, pp2->size, > mpp->size); > mpp->action = ACT_REJECT; > + set_bit_in_array(i, size_mismatch_seen); > } > } > verify_paths(mpp, vecs); > @@ -1119,8 +1129,10 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, > "ignoring" : "removing"); > remove_map(mpp, vecs, 0); > continue; > - } else /* if (r == DOMAP_RETRY) */ > - return r; > + } else /* if (r == DOMAP_RETRY) */ { > + r = 1; > + goto out; > + } > } > if (r == DOMAP_DRY) > continue; > @@ -1161,8 +1173,10 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, > > if (newmp) { > if (mpp->action != ACT_REJECT) { > - if (!vector_alloc_slot(newmp)) > - return 1; > + if (!vector_alloc_slot(newmp)) { > + r = 1; > + goto out; > + } > vector_set_slot(newmp, mpp); > } > else > @@ -1193,7 +1207,10 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, > condlog(2, "%s: remove (dead)", alias); > } > } > - return 0; > + r = 0; > +out: > + free(size_mismatch_seen); > + return r; > } > > struct udev_device *get_udev_device(const char *dev, enum devtypes dev_type) > diff --git a/libmultipath/util.h b/libmultipath/util.h > index a818e29a..1f13c913 100644 > --- a/libmultipath/util.h > +++ b/libmultipath/util.h > @@ -3,6 +3,7 @@ > > #include <sys/types.h> > #include <inttypes.h> > +#include <stdbool.h> > > size_t strchop(char *); > int basenamecpy (const char *src, char *dst, size_t size); > @@ -39,4 +40,19 @@ struct scandir_result { > }; > void free_scandir_result(struct scandir_result *); > > +static inline bool is_bit_set_in_array(unsigned int bit, const uint64_t *arr) > +{ > + return arr[bit / 64] & (1ULL << (bit % 64)) ? 1 : 0; > +} > + > +static inline void set_bit_in_array(unsigned int bit, uint64_t *arr) > +{ > + arr[bit / 64] |= (1ULL << (bit % 64)); > +} > + > +static inline void clear_bit_in_array(unsigned int bit, uint64_t *arr) > +{ > + arr[bit / 64] &= ~(1ULL << (bit % 64)); > +} > + > #endif /* _UTIL_H */ > -- > 2.19.1 -- dm-devel mailing list dm-devel@redhat.com https://www.redhat.com/mailman/listinfo/dm-devel
On Thu, 2018-11-29 at 14:53 -0600, Benjamin Marzinski wrote: > On Wed, Nov 21, 2018 at 11:18:36AM +0100, Martin Wilck wrote: > > When there are paths with the same WWID but different sizes, and > > coalesce_paths() walks the pathvec, it checks paths _after_ > > the current one for size mismatch and sets ACT_REJECT. However, > > these paths will be reached in the main loop later, and this time > > the already handled paths will not be checked for size mismatch; > > thus a map could be created, possibly even with mismatching > > devices. > > > > Fix that by tracking which paths were already discarded, and > > skipping them in the main loop later. > > Previously, multipath would retry if coalesce_paths returned > DOMAP_RETRY. With this patch, DOMAP_RETRY isn't returned, so that no > longer happens. Is this intentional? No, it was an oversight. Thanks for spotting it. Note: it's only relevant for the case where lock_multipath() fails, which should hardly ever happen any more since we are using shared locks (5ec07b3). Whether we still need lock_multipath() is an open question. I personally think we don't. > > Also, I would prefer if coalesce_paths always used named constants > for > the return values, instead of converting them to numbers. OK, I'll send an extra patch for that. Martin
diff --git a/libmultipath/configure.c b/libmultipath/configure.c index 406cd4c9..4ed3cc23 100644 --- a/libmultipath/configure.c +++ b/libmultipath/configure.c @@ -1009,6 +1009,7 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, vector pathvec = vecs->pathvec; struct config *conf; int allow_queueing; + uint64_t *size_mismatch_seen; /* ignore refwwid if it's empty */ if (refwwid && !strlen(refwwid)) @@ -1019,6 +1020,14 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, pp1->mpp = NULL; } } + + if (VECTOR_SIZE(pathvec) == 0) + return 0; + size_mismatch_seen = calloc((VECTOR_SIZE(pathvec) - 1) / 64 + 1, + sizeof(uint64_t)); + if (size_mismatch_seen == NULL) + return 1; + vector_foreach_slot (pathvec, pp1, k) { int invalid; /* skip this path for some reason */ @@ -1038,8 +1047,8 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, continue; } - /* 2. if path already coalesced */ - if (pp1->mpp) + /* 2. if path already coalesced, or seen and discarded */ + if (pp1->mpp || is_bit_set_in_array(k, size_mismatch_seen)) continue; /* 3. if path has disappeared */ @@ -1088,9 +1097,10 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, * ouch, avoid feeding that to the DM */ condlog(0, "%s: size %llu, expected %llu. " - "Discard", pp2->dev_t, pp2->size, + "Discard", pp2->dev, pp2->size, mpp->size); mpp->action = ACT_REJECT; + set_bit_in_array(i, size_mismatch_seen); } } verify_paths(mpp, vecs); @@ -1119,8 +1129,10 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, "ignoring" : "removing"); remove_map(mpp, vecs, 0); continue; - } else /* if (r == DOMAP_RETRY) */ - return r; + } else /* if (r == DOMAP_RETRY) */ { + r = 1; + goto out; + } } if (r == DOMAP_DRY) continue; @@ -1161,8 +1173,10 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, if (newmp) { if (mpp->action != ACT_REJECT) { - if (!vector_alloc_slot(newmp)) - return 1; + if (!vector_alloc_slot(newmp)) { + r = 1; + goto out; + } vector_set_slot(newmp, mpp); } else @@ -1193,7 +1207,10 @@ int coalesce_paths (struct vectors * vecs, vector newmp, char * refwwid, condlog(2, "%s: remove (dead)", alias); } } - return 0; + r = 0; +out: + free(size_mismatch_seen); + return r; } struct udev_device *get_udev_device(const char *dev, enum devtypes dev_type) diff --git a/libmultipath/util.h b/libmultipath/util.h index a818e29a..1f13c913 100644 --- a/libmultipath/util.h +++ b/libmultipath/util.h @@ -3,6 +3,7 @@ #include <sys/types.h> #include <inttypes.h> +#include <stdbool.h> size_t strchop(char *); int basenamecpy (const char *src, char *dst, size_t size); @@ -39,4 +40,19 @@ struct scandir_result { }; void free_scandir_result(struct scandir_result *); +static inline bool is_bit_set_in_array(unsigned int bit, const uint64_t *arr) +{ + return arr[bit / 64] & (1ULL << (bit % 64)) ? 1 : 0; +} + +static inline void set_bit_in_array(unsigned int bit, uint64_t *arr) +{ + arr[bit / 64] |= (1ULL << (bit % 64)); +} + +static inline void clear_bit_in_array(unsigned int bit, uint64_t *arr) +{ + arr[bit / 64] &= ~(1ULL << (bit % 64)); +} + #endif /* _UTIL_H */
When there are paths with the same WWID but different sizes, and coalesce_paths() walks the pathvec, it checks paths _after_ the current one for size mismatch and sets ACT_REJECT. However, these paths will be reached in the main loop later, and this time the already handled paths will not be checked for size mismatch; thus a map could be created, possibly even with mismatching devices. Fix that by tracking which paths were already discarded, and skipping them in the main loop later. Signed-off-by: Martin Wilck <mwilck@suse.com> --- libmultipath/configure.c | 33 +++++++++++++++++++++++++-------- libmultipath/util.h | 16 ++++++++++++++++ 2 files changed, 41 insertions(+), 8 deletions(-)