I found another minor problem:
The old code avoided building the wait table when the timeout is 0, now we
always build the wait table.
Below is an update for the memory allocations within fs/select.c:
* do not build the wait table if timeout==0
* store the first 8 poll_table_entries within the poll_table.
* sys_select: if select is called for <= 256 bits, then store these bits
on the stack.
* sys_poll: replace the arrays with a linked list, and begin that list on
the stack (the first 24 entries).
It means that sys_poll and sys_select both store around 450 bytes on the
stack - IMHO acceptable.
Result: poll() for one file descriptor: down from ~3200 cpu ticks to
~1000 cpu ticks.
The patch was tested with test4-pre4.
Petr, could you double check that I didn't break ncpfs?
-- Manfred<<<<<<<<<< // $Header$ // Kernel Version: // VERSION = 2 // PATCHLEVEL = 4 // SUBLEVEL = 0 // EXTRAVERSION = -test4 --- 2.4/fs/select.c Wed Jul 12 20:06:25 2000 +++ build-2.4/fs/select.c Wed Jul 12 20:54:35 2000 @@ -40,8 +40,15 @@ void poll_freewait(poll_table* pt) { struct poll_table_page * p = pt->table; + struct poll_table_entry * entry; + entry = pt->internal + pt->nr; + while(pt->nr > 0) { + pt->nr--; + entry--; + remove_wait_queue(entry->wait_address,&entry->wait); + fput(entry->filp); + } while (p) { - struct poll_table_entry * entry; struct poll_table_page *old; entry = p->entry + p->nr; @@ -53,41 +60,44 @@ } old = p; p = p->next; - free_page((unsigned long) old); + kfree(old); } } void __pollwait(struct file * filp, wait_queue_head_t * wait_address, poll_table *p) { - struct poll_table_page *table = p->table; - - if (!table || table->nr >= __MAX_POLL_TABLE_ENTRIES) { - struct poll_table_page *new_table; + struct poll_table_entry * entry; - new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL); - if (!new_table) { - p->error = -ENOMEM; - __set_current_state(TASK_RUNNING); - return; - } - new_table->nr = 0; - new_table->entry = (struct poll_table_entry *)(new_table + 1); - new_table->next = table; - p->table = new_table; - table = new_table; - } - - /* Add a new entry */ - { - struct poll_table_entry * entry; + if(p->nr < POLL_TABLE_INTERNAL) { + entry = p->internal + p->nr; + p->nr++; + } else { + struct poll_table_page *table = p->table; + if (!table || table->nr >= __MAX_POLL_TABLE_ENTRIES) { + struct poll_table_page *new_table; + + new_table = (struct poll_table_page *) kmalloc(PAGE_SIZE,GFP_KERNEL); + if (!new_table) { + p->error = -ENOMEM; + __set_current_state(TASK_RUNNING); + return; + } + new_table->nr = 0; + new_table->entry = (struct poll_table_entry *)(new_table + 1); + new_table->next = table; + p->table = new_table; + table = new_table; + } entry = table->entry + table->nr; table->nr++; - get_file(filp); - entry->filp = filp; - entry->wait_address = wait_address; - init_waitqueue_entry(&entry->wait, current); - add_wait_queue(wait_address,&entry->wait); } + + /* Add a new entry */ + get_file(filp); + entry->filp = filp; + entry->wait_address = wait_address; + init_waitqueue_entry(&entry->wait, current); + add_wait_queue(wait_address,&entry->wait); } #define __IN(fds, n) (fds->in + n) @@ -163,7 +173,10 @@ n = retval; poll_initwait(&table); - wait = &table; + if(__timeout) + wait = &table; + else + wait = NULL; retval = 0; for (;;) { set_current_state(TASK_INTERRUPTIBLE); @@ -234,6 +247,7 @@ sys_select(int n, fd_set *inp, fd_set *outp, fd_set *exp, struct timeval *tvp) { fd_set_bits fds; + char inline_bits[192]; char *bits; long timeout; int ret, size; @@ -271,9 +285,13 @@ */ ret = -ENOMEM; size = FDS_BYTES(n); - bits = kmalloc(6 * size, GFP_KERNEL); - if (!bits) - goto out_nofds; + if(size <= sizeof(inline_bits)/6) { + bits = inline_bits; + } else { + bits = kmalloc(6 * size, GFP_KERNEL); + if (!bits) + goto out_nofds; + } fds.in = (unsigned long *) bits; fds.out = (unsigned long *) (bits + size); fds.ex = (unsigned long *) (bits + 2*size); @@ -316,12 +334,18 @@ set_fd_set(n, exp, fds.res_ex); out: - kfree(bits); + if(bits != inline_bits) + kfree(bits); out_nofds: return ret; } -#define POLLFD_PER_PAGE ((PAGE_SIZE) / sizeof(struct pollfd)) +struct poll_list { + struct poll_list *next; + int len; +}; + + #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd)) static void do_pollfd(unsigned int num, struct pollfd * fdpage, poll_table ** pwait, int *count) @@ -355,38 +379,44 @@ } } -static int do_poll(unsigned int nfds, unsigned int nchunks, unsigned int nleft, - struct pollfd *fds[], poll_table *wait, long timeout) + +static int do_poll(int nfds, struct poll_list *list, + poll_table *wait, long timeout) { int count = 0; poll_table* pt = wait; - + for (;;) { - unsigned int i; - + struct poll_list* walk; set_current_state(TASK_INTERRUPTIBLE); - for (i=0; i < nchunks; i++) - do_pollfd(POLLFD_PER_PAGE, fds[i], &pt, &count); - if (nleft) - do_pollfd(nleft, fds[nchunks], &pt, &count); + walk = list; + while(walk != NULL) { + do_pollfd( walk->len, (struct pollfd*)(walk+1), &pt, &count); + walk = walk->next; + } pt = NULL; if (count || !timeout || signal_pending(current)) break; - if(wait->error) { + if(wait->error) return wait->error; - } + timeout = schedule_timeout(timeout); } current->state = TASK_RUNNING; return count; } +#define INLINE_POLL_COUNT 24 asmlinkage long sys_poll(struct pollfd * ufds, unsigned int nfds, long timeout) { - int i, j, fdcount, err; - struct pollfd **fds; + int fdcount, err; + unsigned int i; + struct poll_list *pollwalk; + struct { + struct poll_list head; + struct pollfd entries[INLINE_POLL_COUNT]; + } polllist; poll_table table; - int nchunks, nleft; /* Do a sanity check on nfds ... */ if (nfds > current->files->max_fds) @@ -402,64 +432,65 @@ poll_initwait(&table); err = -ENOMEM; - - fds = NULL; - if (nfds != 0) { - fds = (struct pollfd **)kmalloc( - (1 + (nfds - 1) / POLLFD_PER_PAGE) * sizeof(struct pollfd *), - GFP_KERNEL); - if (fds == NULL) - goto out; - } - - nchunks = 0; - nleft = nfds; - while (nleft > POLLFD_PER_PAGE) { /* allocate complete PAGE_SIZE chunks */ - fds[nchunks] = (struct pollfd *)__get_free_page(GFP_KERNEL); - if (fds[nchunks] == NULL) + polllist.head.next = NULL; + if(nfds <= INLINE_POLL_COUNT) + polllist.head.len = nfds; + else + polllist.head.len = INLINE_POLL_COUNT; + pollwalk = &polllist.head; + i = nfds; + err = -ENOMEM; + goto start; + while(i!=0) { + struct poll_list *pp; + pp = kmalloc(sizeof(struct poll_list)+ + sizeof(struct pollfd)* + (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i), + GFP_KERNEL); + if(pp==NULL) goto out_fds; - nchunks++; - nleft -= POLLFD_PER_PAGE; - } - if (nleft) { /* allocate last PAGE_SIZE chunk, only nleft elements used */ - fds[nchunks] = (struct pollfd *)__get_free_page(GFP_KERNEL); - if (fds[nchunks] == NULL) + pp->next=NULL; + pp->len = (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i); + pollwalk->next = pp; + pollwalk = pp; +start: + if (copy_from_user(pollwalk+1, ufds + nfds-i, + sizeof(struct pollfd)*pollwalk->len)) { + err = -EFAULT; goto out_fds; + } + i -= pollwalk->len; } + + fdcount = do_poll(nfds, &polllist.head, + timeout!=0?&table:NULL, timeout); + /* OK, now copy the revents fields back to user space. */ + i = nfds; + pollwalk = &polllist.head; err = -EFAULT; - for (i=0; i < nchunks; i++) - if (copy_from_user(fds[i], ufds + i*POLLFD_PER_PAGE, PAGE_SIZE)) - goto out_fds1; - if (nleft) { - if (copy_from_user(fds[nchunks], ufds + nchunks*POLLFD_PER_PAGE, - nleft * sizeof(struct pollfd))) - goto out_fds1; + while(pollwalk != NULL) { + struct pollfd * fds = (struct pollfd*)(pollwalk+1); + int j; + + for (j=0; j < pollwalk->len; j++, ufds++) { + if(__put_user(fds[j].revents, &ufds->revents)) + goto out_fds; + } + i -= pollwalk->len; + pollwalk = pollwalk->next; } - - fdcount = do_poll(nfds, nchunks, nleft, fds, &table, timeout); - - /* OK, now copy the revents fields back to user space. */ - for(i=0; i < nchunks; i++) - for (j=0; j < POLLFD_PER_PAGE; j++, ufds++) - __put_user((fds[i] + j)->revents, &ufds->revents); - if (nleft) - for (j=0; j < nleft; j++, ufds++) - __put_user((fds[nchunks] + j)->revents, &ufds->revents); - err = fdcount; if (!fdcount && signal_pending(current)) err = -EINTR; -out_fds1: - if (nleft) - free_page((unsigned long)(fds[nchunks])); out_fds: - for (i=0; i < nchunks; i++) - free_page((unsigned long)(fds[i])); - if (nfds != 0) - kfree(fds); -out: + pollwalk = polllist.head.next; + while(pollwalk!=NULL) { + struct poll_list *pp = pollwalk->next; + kfree(pollwalk); + pollwalk = pp; + } poll_freewait(&table); return err; } --- 2.4/include/linux/poll.h Wed Jul 12 20:06:25 2000 +++ build-2.4/include/linux/poll.h Wed Jul 12 20:16:52 2000 @@ -23,9 +23,12 @@ struct poll_table_entry * entry; }; +#define POLL_TABLE_INTERNAL 8 typedef struct poll_table_struct { int error; + int nr; struct poll_table_page * table; + struct poll_table_entry internal[POLL_TABLE_INTERNAL]; } poll_table; #define __MAX_POLL_TABLE_ENTRIES ((PAGE_SIZE - sizeof (struct poll_table_page)) / sizeof (struct poll_table_entry)) @@ -41,6 +44,7 @@ static inline void poll_initwait(poll_table* pt) { pt->error = 0; + pt->nr = 0; pt->table = NULL; } extern void poll_freewait(poll_table* pt); >>>>>>>>>>
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sat Jul 15 2000 - 21:00:15 EST