557 lines · plain
1.. SPDX-License-Identifier: GPL-2.02 3========4ORANGEFS5========6 7OrangeFS is an LGPL userspace scale-out parallel storage system. It is ideal8for large storage problems faced by HPC, BigData, Streaming Video,9Genomics, Bioinformatics.10 11Orangefs, originally called PVFS, was first developed in 1993 by12Walt Ligon and Eric Blumer as a parallel file system for Parallel13Virtual Machine (PVM) as part of a NASA grant to study the I/O patterns14of parallel programs.15 16Orangefs features include:17 18 * Distributes file data among multiple file servers19 * Supports simultaneous access by multiple clients20 * Stores file data and metadata on servers using local file system21 and access methods22 * Userspace implementation is easy to install and maintain23 * Direct MPI support24 * Stateless25 26 27Mailing List Archives28=====================29 30http://lists.orangefs.org/pipermail/devel_lists.orangefs.org/31 32 33Mailing List Submissions34========================35 36devel@lists.orangefs.org37 38 39Documentation40=============41 42http://www.orangefs.org/documentation/43 44Running ORANGEFS On a Single Server45===================================46 47OrangeFS is usually run in large installations with multiple servers and48clients, but a complete filesystem can be run on a single machine for49development and testing.50 51On Fedora, install orangefs and orangefs-server::52 53 dnf -y install orangefs orangefs-server54 55There is an example server configuration file in56/etc/orangefs/orangefs.conf. Change localhost to your hostname if57necessary.58 59To generate a filesystem to run xfstests against, see below.60 61There is an example client configuration file in /etc/pvfs2tab. It is a62single line. Uncomment it and change the hostname if necessary. This63controls clients which use libpvfs2. This does not control the64pvfs2-client-core.65 66Create the filesystem::67 68 pvfs2-server -f /etc/orangefs/orangefs.conf69 70Start the server::71 72 systemctl start orangefs-server73 74Test the server::75 76 pvfs2-ping -m /pvfsmnt77 78Start the client. The module must be compiled in or loaded before this79point::80 81 systemctl start orangefs-client82 83Mount the filesystem::84 85 mount -t pvfs2 tcp://localhost:3334/orangefs /pvfsmnt86 87Userspace Filesystem Source88===========================89 90http://www.orangefs.org/download91 92Orangefs versions prior to 2.9.3 would not be compatible with the93upstream version of the kernel client.94 95 96Building ORANGEFS on a Single Server97====================================98 99Where OrangeFS cannot be installed from distribution packages, it may be100built from source.101 102You can omit --prefix if you don't care that things are sprinkled around103in /usr/local. As of version 2.9.6, OrangeFS uses Berkeley DB by104default, we will probably be changing the default to LMDB soon.105 106::107 108 ./configure --prefix=/opt/ofs --with-db-backend=lmdb --disable-usrint109 110 make111 112 make install113 114Create an orangefs config file by running pvfs2-genconfig and115specifying a target config file. Pvfs2-genconfig will prompt you116through. Generally it works fine to take the defaults, but you117should use your server's hostname, rather than "localhost" when118it comes to that question::119 120 /opt/ofs/bin/pvfs2-genconfig /etc/pvfs2.conf121 122Create an /etc/pvfs2tab file (localhost is fine)::123 124 echo tcp://localhost:3334/orangefs /pvfsmnt pvfs2 defaults,noauto 0 0 > \125 /etc/pvfs2tab126 127Create the mount point you specified in the tab file if needed::128 129 mkdir /pvfsmnt130 131Bootstrap the server::132 133 /opt/ofs/sbin/pvfs2-server -f /etc/pvfs2.conf134 135Start the server::136 137 /opt/ofs/sbin/pvfs2-server /etc/pvfs2.conf138 139Now the server should be running. Pvfs2-ls is a simple140test to verify that the server is running::141 142 /opt/ofs/bin/pvfs2-ls /pvfsmnt143 144If stuff seems to be working, load the kernel module and145turn on the client core::146 147 /opt/ofs/sbin/pvfs2-client -p /opt/ofs/sbin/pvfs2-client-core148 149Mount your filesystem::150 151 mount -t pvfs2 tcp://`hostname`:3334/orangefs /pvfsmnt152 153 154Running xfstests155================156 157It is useful to use a scratch filesystem with xfstests. This can be158done with only one server.159 160Make a second copy of the FileSystem section in the server configuration161file, which is /etc/orangefs/orangefs.conf. Change the Name to scratch.162Change the ID to something other than the ID of the first FileSystem163section (2 is usually a good choice).164 165Then there are two FileSystem sections: orangefs and scratch.166 167This change should be made before creating the filesystem.168 169::170 171 pvfs2-server -f /etc/orangefs/orangefs.conf172 173To run xfstests, create /etc/xfsqa.config::174 175 TEST_DIR=/orangefs176 TEST_DEV=tcp://localhost:3334/orangefs177 SCRATCH_MNT=/scratch178 SCRATCH_DEV=tcp://localhost:3334/scratch179 180Then xfstests can be run::181 182 ./check -pvfs2183 184 185Options186=======187 188The following mount options are accepted:189 190 acl191 Allow the use of Access Control Lists on files and directories.192 193 intr194 Some operations between the kernel client and the user space195 filesystem can be interruptible, such as changes in debug levels196 and the setting of tunable parameters.197 198 local_lock199 Enable posix locking from the perspective of "this" kernel. The200 default file_operations lock action is to return ENOSYS. Posix201 locking kicks in if the filesystem is mounted with -o local_lock.202 Distributed locking is being worked on for the future.203 204 205Debugging206=========207 208If you want the debug (GOSSIP) statements in a particular209source file (inode.c for example) go to syslog::210 211 echo inode > /sys/kernel/debug/orangefs/kernel-debug212 213No debugging (the default)::214 215 echo none > /sys/kernel/debug/orangefs/kernel-debug216 217Debugging from several source files::218 219 echo inode,dir > /sys/kernel/debug/orangefs/kernel-debug220 221All debugging::222 223 echo all > /sys/kernel/debug/orangefs/kernel-debug224 225Get a list of all debugging keywords::226 227 cat /sys/kernel/debug/orangefs/debug-help228 229 230Protocol between Kernel Module and Userspace231============================================232 233Orangefs is a user space filesystem and an associated kernel module.234We'll just refer to the user space part of Orangefs as "userspace"235from here on out. Orangefs descends from PVFS, and userspace code236still uses PVFS for function and variable names. Userspace typedefs237many of the important structures. Function and variable names in238the kernel module have been transitioned to "orangefs", and The Linux239Coding Style avoids typedefs, so kernel module structures that240correspond to userspace structures are not typedefed.241 242The kernel module implements a pseudo device that userspace243can read from and write to. Userspace can also manipulate the244kernel module through the pseudo device with ioctl.245 246The Bufmap247----------248 249At startup userspace allocates two page-size-aligned (posix_memalign)250mlocked memory buffers, one is used for IO and one is used for readdir251operations. The IO buffer is 41943040 bytes and the readdir buffer is2524194304 bytes. Each buffer contains logical chunks, or partitions, and253a pointer to each buffer is added to its own PVFS_dev_map_desc structure254which also describes its total size, as well as the size and number of255the partitions.256 257A pointer to the IO buffer's PVFS_dev_map_desc structure is sent to a258mapping routine in the kernel module with an ioctl. The structure is259copied from user space to kernel space with copy_from_user and is used260to initialize the kernel module's "bufmap" (struct orangefs_bufmap), which261then contains:262 263 * refcnt264 - a reference counter265 * desc_size - PVFS2_BUFMAP_DEFAULT_DESC_SIZE (4194304) - the IO buffer's266 partition size, which represents the filesystem's block size and267 is used for s_blocksize in super blocks.268 * desc_count - PVFS2_BUFMAP_DEFAULT_DESC_COUNT (10) - the number of269 partitions in the IO buffer.270 * desc_shift - log2(desc_size), used for s_blocksize_bits in super blocks.271 * total_size - the total size of the IO buffer.272 * page_count - the number of 4096 byte pages in the IO buffer.273 * page_array - a pointer to ``page_count * (sizeof(struct page*))`` bytes274 of kcalloced memory. This memory is used as an array of pointers275 to each of the pages in the IO buffer through a call to get_user_pages.276 * desc_array - a pointer to ``desc_count * (sizeof(struct orangefs_bufmap_desc))``277 bytes of kcalloced memory. This memory is further initialized:278 279 user_desc is the kernel's copy of the IO buffer's ORANGEFS_dev_map_desc280 structure. user_desc->ptr points to the IO buffer.281 282 ::283 284 pages_per_desc = bufmap->desc_size / PAGE_SIZE285 offset = 0286 287 bufmap->desc_array[0].page_array = &bufmap->page_array[offset]288 bufmap->desc_array[0].array_count = pages_per_desc = 1024289 bufmap->desc_array[0].uaddr = (user_desc->ptr) + (0 * 1024 * 4096)290 offset += 1024291 .292 .293 .294 bufmap->desc_array[9].page_array = &bufmap->page_array[offset]295 bufmap->desc_array[9].array_count = pages_per_desc = 1024296 bufmap->desc_array[9].uaddr = (user_desc->ptr) +297 (9 * 1024 * 4096)298 offset += 1024299 300 * buffer_index_array - a desc_count sized array of ints, used to301 indicate which of the IO buffer's partitions are available to use.302 * buffer_index_lock - a spinlock to protect buffer_index_array during update.303 * readdir_index_array - a five (ORANGEFS_READDIR_DEFAULT_DESC_COUNT) element304 int array used to indicate which of the readdir buffer's partitions are305 available to use.306 * readdir_index_lock - a spinlock to protect readdir_index_array during307 update.308 309Operations310----------311 312The kernel module builds an "op" (struct orangefs_kernel_op_s) when it313needs to communicate with userspace. Part of the op contains the "upcall"314which expresses the request to userspace. Part of the op eventually315contains the "downcall" which expresses the results of the request.316 317The slab allocator is used to keep a cache of op structures handy.318 319At init time the kernel module defines and initializes a request list320and an in_progress hash table to keep track of all the ops that are321in flight at any given time.322 323Ops are stateful:324 325 * unknown326 - op was just initialized327 * waiting328 - op is on request_list (upward bound)329 * inprogr330 - op is in progress (waiting for downcall)331 * serviced332 - op has matching downcall; ok333 * purged334 - op has to start a timer since client-core335 exited uncleanly before servicing op336 * given up337 - submitter has given up waiting for it338 339When some arbitrary userspace program needs to perform a340filesystem operation on Orangefs (readdir, I/O, create, whatever)341an op structure is initialized and tagged with a distinguishing ID342number. The upcall part of the op is filled out, and the op is343passed to the "service_operation" function.344 345Service_operation changes the op's state to "waiting", puts346it on the request list, and signals the Orangefs file_operations.poll347function through a wait queue. Userspace is polling the pseudo-device348and thus becomes aware of the upcall request that needs to be read.349 350When the Orangefs file_operations.read function is triggered, the351request list is searched for an op that seems ready-to-process.352The op is removed from the request list. The tag from the op and353the filled-out upcall struct are copy_to_user'ed back to userspace.354 355If any of these (and some additional protocol) copy_to_users fail,356the op's state is set to "waiting" and the op is added back to357the request list. Otherwise, the op's state is changed to "in progress",358and the op is hashed on its tag and put onto the end of a list in the359in_progress hash table at the index the tag hashed to.360 361When userspace has assembled the response to the upcall, it362writes the response, which includes the distinguishing tag, back to363the pseudo device in a series of io_vecs. This triggers the Orangefs364file_operations.write_iter function to find the op with the associated365tag and remove it from the in_progress hash table. As long as the op's366state is not "canceled" or "given up", its state is set to "serviced".367The file_operations.write_iter function returns to the waiting vfs,368and back to service_operation through wait_for_matching_downcall.369 370Service operation returns to its caller with the op's downcall371part (the response to the upcall) filled out.372 373The "client-core" is the bridge between the kernel module and374userspace. The client-core is a daemon. The client-core has an375associated watchdog daemon. If the client-core is ever signaled376to die, the watchdog daemon restarts the client-core. Even though377the client-core is restarted "right away", there is a period of378time during such an event that the client-core is dead. A dead client-core379can't be triggered by the Orangefs file_operations.poll function.380Ops that pass through service_operation during a "dead spell" can timeout381on the wait queue and one attempt is made to recycle them. Obviously,382if the client-core stays dead too long, the arbitrary userspace processes383trying to use Orangefs will be negatively affected. Waiting ops384that can't be serviced will be removed from the request list and385have their states set to "given up". In-progress ops that can't386be serviced will be removed from the in_progress hash table and387have their states set to "given up".388 389Readdir and I/O ops are atypical with respect to their payloads.390 391 - readdir ops use the smaller of the two pre-allocated pre-partitioned392 memory buffers. The readdir buffer is only available to userspace.393 The kernel module obtains an index to a free partition before launching394 a readdir op. Userspace deposits the results into the indexed partition395 and then writes them to back to the pvfs device.396 397 - io (read and write) ops use the larger of the two pre-allocated398 pre-partitioned memory buffers. The IO buffer is accessible from399 both userspace and the kernel module. The kernel module obtains an400 index to a free partition before launching an io op. The kernel module401 deposits write data into the indexed partition, to be consumed402 directly by userspace. Userspace deposits the results of read403 requests into the indexed partition, to be consumed directly404 by the kernel module.405 406Responses to kernel requests are all packaged in pvfs2_downcall_t407structs. Besides a few other members, pvfs2_downcall_t contains a408union of structs, each of which is associated with a particular409response type.410 411The several members outside of the union are:412 413 ``int32_t type``414 - type of operation.415 ``int32_t status``416 - return code for the operation.417 ``int64_t trailer_size``418 - 0 unless readdir operation.419 ``char *trailer_buf``420 - initialized to NULL, used during readdir operations.421 422The appropriate member inside the union is filled out for any423particular response.424 425 PVFS2_VFS_OP_FILE_IO426 fill a pvfs2_io_response_t427 428 PVFS2_VFS_OP_LOOKUP429 fill a PVFS_object_kref430 431 PVFS2_VFS_OP_CREATE432 fill a PVFS_object_kref433 434 PVFS2_VFS_OP_SYMLINK435 fill a PVFS_object_kref436 437 PVFS2_VFS_OP_GETATTR438 fill in a PVFS_sys_attr_s (tons of stuff the kernel doesn't need)439 fill in a string with the link target when the object is a symlink.440 441 PVFS2_VFS_OP_MKDIR442 fill a PVFS_object_kref443 444 PVFS2_VFS_OP_STATFS445 fill a pvfs2_statfs_response_t with useless info <g>. It is hard for446 us to know, in a timely fashion, these statistics about our447 distributed network filesystem.448 449 PVFS2_VFS_OP_FS_MOUNT450 fill a pvfs2_fs_mount_response_t which is just like a PVFS_object_kref451 except its members are in a different order and "__pad1" is replaced452 with "id".453 454 PVFS2_VFS_OP_GETXATTR455 fill a pvfs2_getxattr_response_t456 457 PVFS2_VFS_OP_LISTXATTR458 fill a pvfs2_listxattr_response_t459 460 PVFS2_VFS_OP_PARAM461 fill a pvfs2_param_response_t462 463 PVFS2_VFS_OP_PERF_COUNT464 fill a pvfs2_perf_count_response_t465 466 PVFS2_VFS_OP_FSKEY467 file a pvfs2_fs_key_response_t468 469 PVFS2_VFS_OP_READDIR470 jamb everything needed to represent a pvfs2_readdir_response_t into471 the readdir buffer descriptor specified in the upcall.472 473Userspace uses writev() on /dev/pvfs2-req to pass responses to the requests474made by the kernel side.475 476A buffer_list containing:477 478 - a pointer to the prepared response to the request from the479 kernel (struct pvfs2_downcall_t).480 - and also, in the case of a readdir request, a pointer to a481 buffer containing descriptors for the objects in the target482 directory.483 484... is sent to the function (PINT_dev_write_list) which performs485the writev.486 487PINT_dev_write_list has a local iovec array: struct iovec io_array[10];488 489The first four elements of io_array are initialized like this for all490responses::491 492 io_array[0].iov_base = address of local variable "proto_ver" (int32_t)493 io_array[0].iov_len = sizeof(int32_t)494 495 io_array[1].iov_base = address of global variable "pdev_magic" (int32_t)496 io_array[1].iov_len = sizeof(int32_t)497 498 io_array[2].iov_base = address of parameter "tag" (PVFS_id_gen_t)499 io_array[2].iov_len = sizeof(int64_t)500 501 io_array[3].iov_base = address of out_downcall member (pvfs2_downcall_t)502 of global variable vfs_request (vfs_request_t)503 io_array[3].iov_len = sizeof(pvfs2_downcall_t)504 505Readdir responses initialize the fifth element io_array like this::506 507 io_array[4].iov_base = contents of member trailer_buf (char *)508 from out_downcall member of global variable509 vfs_request510 io_array[4].iov_len = contents of member trailer_size (PVFS_size)511 from out_downcall member of global variable512 vfs_request513 514Orangefs exploits the dcache in order to avoid sending redundant515requests to userspace. We keep object inode attributes up-to-date with516orangefs_inode_getattr. Orangefs_inode_getattr uses two arguments to517help it decide whether or not to update an inode: "new" and "bypass".518Orangefs keeps private data in an object's inode that includes a short519timeout value, getattr_time, which allows any iteration of520orangefs_inode_getattr to know how long it has been since the inode was521updated. When the object is not new (new == 0) and the bypass flag is not522set (bypass == 0) orangefs_inode_getattr returns without updating the inode523if getattr_time has not timed out. Getattr_time is updated each time the524inode is updated.525 526Creation of a new object (file, dir, sym-link) includes the evaluation of527its pathname, resulting in a negative directory entry for the object.528A new inode is allocated and associated with the dentry, turning it from529a negative dentry into a "productive full member of society". Orangefs530obtains the new inode from Linux with new_inode() and associates531the inode with the dentry by sending the pair back to Linux with532d_instantiate().533 534The evaluation of a pathname for an object resolves to its corresponding535dentry. If there is no corresponding dentry, one is created for it in536the dcache. Whenever a dentry is modified or verified Orangefs stores a537short timeout value in the dentry's d_time, and the dentry will be trusted538for that amount of time. Orangefs is a network filesystem, and objects539can potentially change out-of-band with any particular Orangefs kernel module540instance, so trusting a dentry is risky. The alternative to trusting541dentries is to always obtain the needed information from userspace - at542least a trip to the client-core, maybe to the servers. Obtaining information543from a dentry is cheap, obtaining it from userspace is relatively expensive,544hence the motivation to use the dentry when possible.545 546The timeout values d_time and getattr_time are jiffy based, and the547code is designed to avoid the jiffy-wrap problem::548 549 "In general, if the clock may have wrapped around more than once, there550 is no way to tell how much time has elapsed. However, if the times t1551 and t2 are known to be fairly close, we can reliably compute the552 difference in a way that takes into account the possibility that the553 clock may have wrapped between times."554 555from course notes by instructor Andy Wang556 557