linux-dev/fs/afs/server.c, branch linus/master

afs: Fix mmap coherency vs 3rd-party changes

2021-09-13T08:10:39Z

Fix the coherency management of mmap'd data such that 3rd-party changes become visible as soon as possible after the callback notification is delivered by the fileserver. This is done by the following means: (1) When we break a callback on a vnode specified by the CB.CallBack call from the server, we queue a work item (vnode->cb_work) to go and clobber all the PTEs mapping to that inode. This causes the CPU to trip through the ->map_pages() and ->page_mkwrite() handlers if userspace attempts to access the page(s) again. (Ideally, this would be done in the service handler for CB.CallBack, but the server is waiting for our reply before considering, and we have a list of vnodes, all of which need breaking - and the process of getting the mmap_lock and stripping the PTEs on all CPUs could be quite slow.) (2) Call afs_validate() from the ->map_pages() handler to check to see if the file has changed and to get a new callback promise from the server. Also handle the fileserver telling us that it's dropping all callbacks, possibly after it's been restarted by sending us a CB.InitCallBackState* call by the following means: (3) Maintain a per-cell list of afs files that are currently mmap'd (cell->fs_open_mmaps). (4) Add a work item to each server that is invoked if there are any open mmaps when CB.InitCallBackState happens. This work item goes through the aforementioned list and invokes the vnode->cb_work work item for each one that is currently using this server. This causes the PTEs to be cleared, causing ->map_pages() or ->page_mkwrite() to be called again, thereby calling afs_validate() again. I've chosen to simply strip the PTEs at the point of notification reception rather than invalidate all the pages as well because (a) it's faster, (b) we may get a notification for other reasons than the data being altered (in which case we don't want to clobber the pagecache) and (c) we need to ask the server to find out - and I don't want to wait for the reply before holding up userspace. This was tested using the attached test program: #include #include #include #include #include #include int main(int argc, char *argv[]) { size_t size = getpagesize(); unsigned char *p; bool mod = (argc == 3); int fd; if (argc != 2 && argc != 3) { fprintf(stderr, "Format: %s [mod]\n", argv[0]); exit(2); } fd = open(argv[1], mod ? O_RDWR : O_RDONLY); if (fd < 0) { perror(argv[1]); exit(1); } p = mmap(NULL, size, mod ? PROT_READ|PROT_WRITE : PROT_READ, MAP_SHARED, fd, 0); if (p == MAP_FAILED) { perror("mmap"); exit(1); } for (;;) { if (mod) { p[0]++; msync(p, size, MS_ASYNC); fsync(fd); } printf("%02x", p[0]); fflush(stdout); sleep(1); } } It runs in two modes: in one mode, it mmaps a file, then sits in a loop reading the first byte, printing it and sleeping for a second; in the second mode it mmaps a file, then sits in a loop incrementing the first byte and flushing, then printing and sleeping. Two instances of this program can be run on different machines, one doing the reading and one doing the writing. The reader should see the changes made by the writer, but without this patch, they aren't because validity checking is being done lazily - only on entry to the filesystem. Testing the InitCallBackState change is more complicated. The server has to be taken offline, the saved callback state file removed and then the server restarted whilst the reading-mode program continues to run. The client machine then has to poke the server to trigger the InitCallBackState call. Signed-off-by: David Howells Tested-by: Markus Suvanto cc: linux-afs@lists.infradead.org Link: https://lore.kernel.org/r/163111668833.283156.382633263709075739.stgit@warthog.procyon.org.uk/

afs: Don't assert on unpurgeable server records

2020-10-16T13:39:34Z

Don't give an assertion failure on unpurgeable afs_server records - which kills the thread - but rather emit a trace line when we are purging a record (which only happens during network namespace removal or rmmod) and print a notice of the problem. Signed-off-by: David Howells

afs: Fix hang on rmmod due to outstanding timer

2020-06-20T19:01:58Z

The fileserver probe timer, net->fs_probe_timer, isn't cancelled when the kafs module is being removed and so the count it holds on net->servers_outstanding doesn't get dropped.. This causes rmmod to wait forever. The hung process shows a stack like: afs_purge_servers+0x1b5/0x23c [kafs] afs_net_exit+0x44/0x6e [kafs] ops_exit_list+0x72/0x93 unregister_pernet_operations+0x14c/0x1ba unregister_pernet_subsys+0x1d/0x2a afs_exit+0x29/0x6f [kafs] __do_sys_delete_module.isra.0+0x1a2/0x24b do_syscall_64+0x51/0x95 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Fix this by: (1) Attempting to cancel the probe timer and, if successful, drop the count that the timer was holding. (2) Make the timer function just drop the count and not schedule the prober if the afs portion of net namespace is being destroyed. Also, whilst we're at it, make the following changes: (3) Initialise net->servers_outstanding to 1 and decrement it before waiting on it so that it doesn't generate wake up events by being decremented to 0 until we're cleaning up. (4) Switch the atomic_dec() on ->servers_outstanding for ->fs_timer in afs_purge_servers() to use the helper function for that. Fixes: f6cbb368bcb0 ("afs: Actively poll fileservers to maintain NAT or firewall openings") Signed-off-by: David Howells Signed-off-by: Linus Torvalds

afs: Don't use probe running state to make decisions outside probe code

2020-06-04T14:37:58Z

Don't use the running state for fileserver probes to make decisions about which server to use as the state is cleared at the start of a probe and also intermediate values might be misleading. Instead, add a separate 'latest known' rtt in the afs_server struct and a flag to indicate if the server is known to be responding and update these as and when we know what to change them to. Signed-off-by: David Howells

afs: Fix the by-UUID server tree to allow servers with the same UUID

2020-06-04T14:37:57Z

Whilst it shouldn't happen, it is possible for multiple fileservers to share a UUID, particularly if an entire cell has been duplicated, UUIDs and all. In such a case, it's not necessarily possible to map the effect of the CB.InitCallBackState3 incoming RPC to a specific server unambiguously by UUID and thus to a specific cell. Indeed, there's a problem whereby multiple server records may need to occupy the same spot in the rb_tree rooted in the afs_net struct. Fix this by allowing servers to form a list, with the head of the list in the tree. When the front entry in the list is removed, the second in the list just replaces it. afs_init_callback_state() then just goes down the line, poking each server in the list. This means that some servers will be unnecessarily poked, unfortunately. An alternative would be to route by call parameters. Reported-by: Jeffrey Altman Signed-off-by: David Howells Fixes: d2ddc776a458 ("afs: Overhaul volume and server record caching and fileserver rotation")

afs: Reorganise volume and server trees to be rooted on the cell

2020-06-04T14:37:57Z

Reorganise afs_volume objects such that they're in a tree keyed on volume ID, rooted at on an afs_cell object rather than being in multiple trees, each of which is rooted on an afs_server object. afs_server structs become per-cell and acquire a pointer to the cell. The process of breaking a callback then starts with finding the server by its network address, following that to the cell and then looking up each volume ID in the volume tree. This is simpler than the afs_vol_interest/afs_cb_interest N:M mapping web and allows those structs and the code for maintaining them to be simplified or removed. It does make a couple of things a bit more tricky, though: (1) Operations now start with a volume, not a server, so there can be more than one answer as to whether or not the server we'll end up using supports the FS.InlineBulkStatus RPC. (2) CB RPC operations that specify the server UUID. There's still a tree of servers by UUID on the afs_net struct, but the UUIDs in it aren't guaranteed unique. Signed-off-by: David Howells

afs: Build an abstraction around an "operation" concept

2020-06-04T14:37:17Z

Turn the afs_operation struct into the main way that most fileserver operations are managed. Various things are added to the struct, including the following: (1) All the parameters and results of the relevant operations are moved into it, removing corresponding fields from the afs_call struct. afs_call gets a pointer to the op. (2) The target volume is made the main focus of the operation, rather than the target vnode(s), and a bunch of op->vnode->volume are made op->volume instead. (3) Two vnode records are defined (op->file[]) for the vnode(s) involved in most operations. The vnode record (struct afs_vnode_param) contains: - The vnode pointer. - The fid of the vnode to be included in the parameters or that was returned in the reply (eg. FS.MakeDir). - The status and callback information that may be returned in the reply about the vnode. - Callback break and data version tracking for detecting simultaneous third-parth changes. (4) Pointers to dentries to be updated with new inodes. (5) An operations table pointer. The table includes pointers to functions for issuing AFS and YFS-variant RPCs, handling the success and abort of an operation and handling post-I/O-lock local editing of a directory. To make this work, the following function restructuring is made: (A) The rotation loop that issues calls to fileservers that can be found in each function that wants to issue an RPC (such as afs_mkdir()) is extracted out into common code, in a new file called fs_operation.c. (B) The rotation loops, such as the one in afs_mkdir(), are replaced with a much smaller piece of code that allocates an operation, sets the parameters and then calls out to the common code to do the actual work. (C) The code for handling the success and failure of an operation are moved into operation functions (as (5) above) and these are called from the core code at appropriate times. (D) The pseudo inode getting stuff used by the dynamic root code is moved over into dynroot.c. (E) struct afs_iget_data is absorbed into the operation struct and afs_iget() expects to be given an op pointer and a vnode record. (F) Point (E) doesn't work for the root dir of a volume, but we know the FID in advance (it's always vnode 1, unique 1), so a separate inode getter, afs_root_iget(), is provided to special-case that. (G) The inode status init/update functions now also take an op and a vnode record. (H) The RPC marshalling functions now, for the most part, just take an afs_operation struct as their only argument. All the data they need is held there. The result delivery functions write their answers there as well. (I) The call is attached to the operation and then the operation core does the waiting. And then the new operation code is, for the moment, made to just initialise the operation, get the appropriate vnode I/O locks and do the same rotation loop as before. This lays the foundation for the following changes in the future: (*) Overhauling the rotation (again). (*) Support for asynchronous I/O, where the fileserver rotation must be done asynchronously also. Signed-off-by: David Howells

afs: Rename struct afs_fs_cursor to afs_operation

2020-05-31T14:19:52Z

As a prelude to implementing asynchronous fileserver operations in the afs filesystem, rename struct afs_fs_cursor to afs_operation. This struct is going to form the core of the operation management and is going to acquire more members in later. Signed-off-by: David Howells

afs: Make callback processing more efficient.

2020-05-31T14:19:51Z

afs_vol_interest objects represent the volume IDs currently being accessed from a fileserver. These hold lists of afs_cb_interest objects that repesent the superblocks using that volume ID on that server. When a callback notification from the server telling of a modification by another client arrives, the volume ID specified in the notification is looked up in the server's afs_vol_interest list. Through the afs_cb_interest list, the relevant superblocks can be iterated over and the specific inode looked up and marked in each one. Make the following efficiency improvements: (1) Hold rcu_read_lock() over the entire processing rather than locking it each time. (2) Do all the callbacks for each vid together rather than individually. Each volume then only needs to be looked up once. (3) afs_vol_interest objects are now stored in an rb_tree rather than a flat list to reduce the lookup step count. (4) afs_vol_interest lookup is now done with RCU, but because it's in an rb_tree which may rotate under us, a seqlock is used so that if it changes during the walk, we repeat the walk with a lock held. With this and the preceding patch which adds RCU-based lookups in the inode cache, target volumes/vnodes can be taken without the need to take any locks, except on the target itself. Signed-off-by: David Howells

afs: Actively poll fileservers to maintain NAT or firewall openings

2020-05-31T14:19:51Z

When an AFS client accesses a file, it receives a limited-duration callback promise that the server will notify it if another client changes a file. This callback duration can be a few hours in length. If a client mounts a volume and then an application prevents it from being unmounted, say by chdir'ing into it, but then does nothing for some time, the rxrpc_peer record will expire and rxrpc-level keepalive will cease. If there is NAT or a firewall between the client and the server, the route back for the server may close after a comparatively short duration, meaning that attempts by the server to notify the client may then bounce. The client, however, may (so far as it knows) still have a valid unexpired promise and will then rely on its cached data and will not see changes made on the server by a third party until it incidentally rechecks the status or the promise needs renewal. To deal with this, the client needs to regularly probe the server. This has two effects: firstly, it keeps a route open back for the server, and secondly, it causes the server to disgorge any notifications that got queued up because they couldn't be sent. Fix this by adding a mechanism to emit regular probes. Two levels of probing are made available: Under normal circumstances the 'slow' queue will be used for a fileserver - this just probes the preferred address once every 5 mins or so; however, if server fails to respond to any probes, the server will shift to the 'fast' queue from which all its interfaces will be probed every 30s. When it finally responds, the record will switch back to the slow queue. Further notes: (1) Probing is now no longer driven from the fileserver rotation algorithm. (2) Probes are dispatched to all interfaces on a fileserver when that an afs_server object is set up to record it. (3) The afs_server object is removed from the probe queues when we start to probe it. afs_is_probing_server() returns true if it's not listed - ie. it's undergoing probing. (4) The afs_server object is added back on to the probe queue when the final outstanding probe completes, but the probed_at time is set when we're about to launch a probe so that it's not dependent on the probe duration. (5) The timer and the work item added for this must be handed a count on net->servers_outstanding, which they hand on or release. This makes sure that network namespace cleanup waits for them. Fixes: d2ddc776a458 ("afs: Overhaul volume and server record caching and fileserver rotation") Reported-by: Dave Botsch Signed-off-by: David Howells