mp_fget.c

这是国外的resip协议栈

		extending = ret = 0;
		R_LOCK(dbenv, dbmp->reginfo);
		switch (flags) {
		case DB_MPOOL_NEW:
			extending = 1;
			if (mfp->maxpgno != 0 &&
			    mfp->last_pgno >= mfp->maxpgno) {
				__db_err(dbenv, "%s: file limited to %lu pages",
				    __memp_fn(dbmfp), (u_long)mfp->maxpgno);
				ret = ENOSPC;
			} else
				*pgnoaddr = mfp->last_pgno + 1;
			break;
		case DB_MPOOL_CREATE:
			if (mfp->maxpgno != 0 && *pgnoaddr > mfp->maxpgno) {
				__db_err(dbenv, "%s: file limited to %lu pages",
				    __memp_fn(dbmfp), (u_long)mfp->maxpgno);
				ret = ENOSPC;
			} else
				extending = *pgnoaddr > mfp->last_pgno;
			break;
		default:
			ret = *pgnoaddr > mfp->last_pgno ? DB_PAGE_NOTFOUND : 0;
			break;
		}
		R_UNLOCK(dbenv, dbmp->reginfo);
		if (ret != 0)
			goto err;

		/*
		 * !!!
		 * In the DB_MPOOL_NEW code path, mf_offset and n_cache have
		 * not yet been initialized.
		 */
		mf_offset = R_OFFSET(dbmp->reginfo, mfp);
		n_cache = NCACHE(mp, mf_offset, *pgnoaddr);
		c_mp = dbmp->reginfo[n_cache].primary;

		/* Allocate a new buffer header and data space. */
		if ((ret = __memp_alloc(dbmp,
		    &dbmp->reginfo[n_cache], mfp, 0, NULL, &alloc_bhp)) != 0)
			goto err;
#ifdef DIAGNOSTIC
		if ((uintptr_t)alloc_bhp->buf & (sizeof(size_t) - 1)) {
			__db_err(dbenv,
		    "DB_MPOOLFILE->get: buffer data is NOT size_t aligned");
			ret = __db_panic(dbenv, EINVAL);
			goto err;
		}
#endif
		/*
		 * If we are extending the file, we'll need the region lock
		 * again.
		 */
		if (extending)
			R_LOCK(dbenv, dbmp->reginfo);

		/*
		 * DB_MPOOL_NEW does not guarantee you a page unreferenced by
		 * any other thread of control.  (That guarantee is interesting
		 * for DB_MPOOL_NEW, unlike DB_MPOOL_CREATE, because the caller
		 * did not specify the page number, and so, may reasonably not
		 * have any way to lock the page outside of mpool.) Regardless,
		 * if we allocate the page, and some other thread of control
		 * requests the page by number, we will not detect that and the
		 * thread of control that allocated using DB_MPOOL_NEW may not
		 * have a chance to initialize the page.  (Note: we *could*
		 * detect this case if we set a flag in the buffer header which
		 * guaranteed that no gets of the page would succeed until the
		 * reference count went to 0, that is, until the creating page
		 * put the page.)  What we do guarantee is that if two threads
		 * of control are both doing DB_MPOOL_NEW calls, they won't
		 * collide, that is, they won't both get the same page.
		 *
		 * There's a possibility that another thread allocated the page
		 * we were planning to allocate while we were off doing buffer
		 * allocation.  We can do that by making sure the page number
		 * we were going to use is still available.  If it's not, then
		 * we check to see if the next available page number hashes to
		 * the same mpool region as the old one -- if it does, we can
		 * continue, otherwise, we have to start over.
		 */
		if (flags == DB_MPOOL_NEW && *pgnoaddr != mfp->last_pgno + 1) {
			*pgnoaddr = mfp->last_pgno + 1;
			if (n_cache != NCACHE(mp, mf_offset, *pgnoaddr)) {
				/*
				 * flags == DB_MPOOL_NEW, so extending is set
				 * and we're holding the region locked.
				 */
				R_UNLOCK(dbenv, dbmp->reginfo);

				R_LOCK(dbenv, &dbmp->reginfo[n_cache]);
				__db_shalloc_free(
				    &dbmp->reginfo[n_cache], alloc_bhp);
				c_mp->stat.st_pages--;
				R_UNLOCK(dbenv, &dbmp->reginfo[n_cache]);

				alloc_bhp = NULL;
				goto alloc;
			}
		}

		/*
		 * We released the region lock, so another thread might have
		 * extended the file.  Update the last_pgno and initialize
		 * the file, as necessary, if we extended the file.
		 */
		if (extending) {
			if (*pgnoaddr > mfp->last_pgno)
				mfp->last_pgno = *pgnoaddr;

			R_UNLOCK(dbenv, dbmp->reginfo);
			if (ret != 0)
				goto err;
		}
		goto hb_search;
	case SECOND_FOUND:
		/*
		 * We allocated buffer space for the requested page, but then
		 * found the page in the buffer cache on our second check.
		 * That's OK -- we can use the page we found in the pool,
		 * unless DB_MPOOL_NEW is set.
		 *
		 * Free the allocated memory, we no longer need it.  Since we
		 * can't acquire the region lock while holding the hash bucket
		 * lock, we have to release the hash bucket and re-acquire it.
		 * That's OK, because we have the buffer pinned down.
		 */
		MUTEX_UNLOCK(dbenv, &hp->hash_mutex);
		R_LOCK(dbenv, &dbmp->reginfo[n_cache]);
		__db_shalloc_free(&dbmp->reginfo[n_cache], alloc_bhp);
		c_mp->stat.st_pages--;
		alloc_bhp = NULL;
		R_UNLOCK(dbenv, &dbmp->reginfo[n_cache]);

		/*
		 * We can't use the page we found in the pool if DB_MPOOL_NEW
		 * was set.  (For details, see the above comment beginning
		 * "DB_MPOOL_NEW does not guarantee you a page unreferenced by
		 * any other thread of control".)  If DB_MPOOL_NEW is set, we
		 * release our pin on this particular buffer, and try to get
		 * another one.
		 */
		if (flags == DB_MPOOL_NEW) {
			--bhp->ref;
			b_incr = 0;
			goto alloc;
		}

		/* We can use the page -- get the bucket lock. */
		MUTEX_LOCK(dbenv, &hp->hash_mutex);
		break;
	case SECOND_MISS:
		/*
		 * We allocated buffer space for the requested page, and found
		 * the page still missing on our second pass through the buffer
		 * cache.  Instantiate the page.
		 */
		bhp = alloc_bhp;
		alloc_bhp = NULL;

		/*
		 * Initialize all the BH and hash bucket fields so we can call
		 * __memp_bhfree if an error occurs.
		 *
		 * Append the buffer to the tail of the bucket list and update
		 * the hash bucket's priority.
		 */
		b_incr = 1;

		/*lint --e{668} (flexelint: bhp cannot be NULL). */
		memset(bhp, 0, sizeof(BH));
		bhp->ref = 1;
		bhp->priority = UINT32_MAX;
		bhp->pgno = *pgnoaddr;
		bhp->mf_offset = mf_offset;
		SH_TAILQ_INSERT_TAIL(&hp->hash_bucket, bhp, hq);
		hp->hash_priority =
		    SH_TAILQ_FIRST(&hp->hash_bucket, __bh)->priority;

		/* If we extended the file, make sure the page is never lost. */
		if (extending) {
			++hp->hash_page_dirty;
			F_SET(bhp, BH_DIRTY | BH_DIRTY_CREATE);
		}

		/*
		 * If we created the page, zero it out.  If we didn't create
		 * the page, read from the backing file.
		 *
		 * !!!
		 * DB_MPOOL_NEW doesn't call the pgin function.
		 *
		 * If DB_MPOOL_CREATE is used, then the application's pgin
		 * function has to be able to handle pages of 0's -- if it
		 * uses DB_MPOOL_NEW, it can detect all of its page creates,
		 * and not bother.
		 *
		 * If we're running in diagnostic mode, smash any bytes on the
		 * page that are unknown quantities for the caller.
		 *
		 * Otherwise, read the page into memory, optionally creating it
		 * if DB_MPOOL_CREATE is set.
		 */
		if (extending) {
			if (mfp->clear_len == 0)
				memset(bhp->buf, 0, mfp->stat.st_pagesize);
			else {
				memset(bhp->buf, 0, mfp->clear_len);
#if defined(DIAGNOSTIC) || defined(UMRW)
				memset(bhp->buf + mfp->clear_len, CLEAR_BYTE,
				    mfp->stat.st_pagesize - mfp->clear_len);
#endif
			}

			if (flags == DB_MPOOL_CREATE && mfp->ftype != 0)
				F_SET(bhp, BH_CALLPGIN);

			++mfp->stat.st_page_create;
		} else {
			F_SET(bhp, BH_TRASH);
			++mfp->stat.st_cache_miss;
		}

		/* Increment buffer count referenced by MPOOLFILE. */
		MUTEX_LOCK(dbenv, &mfp->mutex);
		++mfp->block_cnt;
		MUTEX_UNLOCK(dbenv, &mfp->mutex);

		/*
		 * Initialize the mutex.  This is the last initialization step,
		 * because it's the only one that can fail, and everything else
		 * must be set up or we can't jump to the err label because it
		 * will call __memp_bhfree.
		 */
		if ((ret = __db_mutex_setup(dbenv,
		    &dbmp->reginfo[n_cache], &bhp->mutex, 0)) != 0)
			goto err;
	}

	DB_ASSERT(bhp->ref != 0);

	/*
	 * If we're the only reference, update buffer and bucket priorities.
	 * We may be about to release the hash bucket lock, and everything
	 * should be correct, first.  (We've already done this if we created
	 * the buffer, so there is no need to do it again.)
	 */
	if (state != SECOND_MISS && bhp->ref == 1) {
		bhp->priority = UINT32_MAX;
		SH_TAILQ_REMOVE(&hp->hash_bucket, bhp, hq, __bh);
		SH_TAILQ_INSERT_TAIL(&hp->hash_bucket, bhp, hq);
		hp->hash_priority =
		    SH_TAILQ_FIRST(&hp->hash_bucket, __bh)->priority;
	}

	/*
	 * BH_TRASH --
	 * The buffer we found may need to be filled from the disk.
	 *
	 * It's possible for the read function to fail, which means we fail as
	 * well.  Note, the __memp_pgread() function discards and reacquires
	 * the hash lock, so the buffer must be pinned down so that it cannot
	 * move and its contents are unchanged.  Discard the buffer on failure
	 * unless another thread is waiting on our I/O to complete.  It's OK to
	 * leave the buffer around, as the waiting thread will see the BH_TRASH
	 * flag set, and will also attempt to discard it.  If there's a waiter,
	 * we need to decrement our reference count.
	 */
	if (F_ISSET(bhp, BH_TRASH) &&
	    (ret = __memp_pgread(dbmfp,
	    &hp->hash_mutex, bhp, LF_ISSET(DB_MPOOL_CREATE) ? 1 : 0)) != 0)
		goto err;

	/*
	 * BH_CALLPGIN --
	 * The buffer was processed for being written to disk, and now has
	 * to be re-converted for use.
	 */
	if (F_ISSET(bhp, BH_CALLPGIN)) {
		if ((ret = __memp_pg(dbmfp, bhp, 1)) != 0)
			goto err;
		F_CLR(bhp, BH_CALLPGIN);
	}

	MUTEX_UNLOCK(dbenv, &hp->hash_mutex);

#ifdef DIAGNOSTIC
	/* Update the file's pinned reference count. */
	R_LOCK(dbenv, dbmp->reginfo);
	++dbmfp->pinref;
	R_UNLOCK(dbenv, dbmp->reginfo);

	/*
	 * We want to switch threads as often as possible, and at awkward
	 * times.  Yield every time we get a new page to ensure contention.
	 */
	if (F_ISSET(dbenv, DB_ENV_YIELDCPU))
		__os_yield(dbenv, 1);
#endif

	*(void **)addrp = bhp->buf;
	return (0);

err:	/*
	 * Discard our reference.  If we're the only reference, discard the
	 * the buffer entirely.  If we held a reference to a buffer, we are
	 * also still holding the hash bucket mutex.
	 */
	if (b_incr) {
		if (bhp->ref == 1)
			__memp_bhfree(dbmp, hp, bhp, BH_FREE_FREEMEM);
		else {
			--bhp->ref;
			MUTEX_UNLOCK(dbenv, &hp->hash_mutex);
		}
	}

	/* If alloc_bhp is set, free the memory. */
	if (alloc_bhp != NULL) {
		R_LOCK(dbenv, &dbmp->reginfo[n_cache]);
		__db_shalloc_free(&dbmp->reginfo[n_cache], alloc_bhp);
		c_mp->stat.st_pages--;
		R_UNLOCK(dbenv, &dbmp->reginfo[n_cache]);
	}

	return (ret);
}