Don't ask for a "test" - ask for a "model" predicting the values of one of the variables (e.g. "blood pressure") as a function of the values of the other variable (e.g. "Vit.D"). This prediction can be adjusted by other variables. You will not only get a "degree of association" (like Spearman's rho) that are usually difficult to interpret but you will get (i.e. estimate) a functional relationship between these variable, you will get effect sizes (e.g. "how much does the BP change per µg/dl additional vit.D?"), and you will get prediction intervals (e.g. "what BP can I expect from a person with x µg/dl Vit.D?").

The keyword you need to search/google is "linear model".

You may also want to enhance your analysis, along the lines suggested above by Dr. Wilhem by trying to account for variables that are directly related to circulating vitamin D such as seasonal variation (sampling dates) and type of clothing of the involved patients: as an extreme example, religious, heavily covered women would have much less sun exposure than modern women using more westernized type of clothing.

I also agree with Jochen. Spearman's Rho or Pearson's r (if you achieve a normal distribution) is good for simple correlation between only 2 variables. But when you start to factor in more variables you need to start thinking about linear regression. Linear instead of logistic in this case because your outcome variable (Blood Pressure) I think is continuous. If you choose to change your outcome variable into categorical (i.e. hypertensive or not) then you may need logistic regression. which one you choose (logistic or linear) will depend on your data. and you may need a bigger sample to achieve normal enough distribution and find the relationship between the  variables.

Jochen is correct concerning the notion of a model.

If both Vit. D and blood pressure are measures it is Pearson Correlation you would use, hence the notion of a linear model (also this could then further involve linear regression).

The other variables you mention (Gender, BMI) will result in different tests (due to the nature of their data types) this will /can enhance an overall model, but in some cases yield weaker less impactive statistics (Spearman's Correlation and Chi-square - tests you may use here - carry less impact than those linear model suggestions)

Responding to Robert:

The relationship between the predictors and the response does NOT need to be linear per se. Their relation should only be linear in the underlying statistical model, where the variables may be transformed to linearize the relationships. So linear models can well be used to model "curved relationships" as well. 

However, the term "linear model" does not even aim at the shape of the relationship but at the functional form of the predictor function: this function is "linear in its coefficients" (the coefficients are the effect sized that are to be estimated based on the available data and the assumed probability model).The predictor function of a linear model is of the general form

y = b0 + b1*f1(x1) + b2*f2(x2) + ... + bn*fn(xn)

Where the b's are the coefficients, the f's are some functions (identity, log, square, sin, etc.) and the x's are the predictor values. There can further be terms that contain products of other predictors, like bk*fa(xa)*fb(xb), what is called an interaction (quite often this is the scientifically most interesting part!).

Thus,

y = b0 + b1*log(x1) + b2*x2²

is a linear model (where the relation of both predictors and the response are non-linear!), but

y = b0 + b1*log(b2+x1)

is not (there is no way to get b1 out of the log).

If the domain of y is limited (like y>0 for counts or concentrations, or 0

Dear Ebrahim!

Other friends added very good answers! But I suggest you read about Linear Models by John Neter and Generalized Linear Models (GLM) in books like:

Categorical Data Analysis for the Behavioral and Social Sciences - by  Razia Azen

or Categorical Data Analysis by Agressti.